# DataArc > DataArc (Dataarc Inc., founded 2025, Shenzhen, China) builds speech recognition (ASR), text-to-speech (TTS), 8-second voice cloning, AI meeting notes (ArcNotes), AI video dubbing with lip sync (ArcDub), and multilingual subtitles focused on Thai, Chinese, and English—available via free browser tools and developer APIs. ## Key Facts & Service Limits - Canonical Website: https://www.dataarc.tech - Legal Entity: Dataarc Inc. (Founded 2025, Shenzhen, China) - Business Contact: bd@dataarctech.com - Supported Languages: Thai (th), Chinese (zh-CN), English (en), plus Thai-Chinese-English intra-sentence code-switching; multilingual subtitles demo supports 8 languages (ja, fr, de, th, zh-CN, ko, en-US, es). - Free Browser Tier: 5 free minutes/day without sign-up (30 min/day signed in), resetting daily at 00:00 UTC+8; up to 3 min/file in the browser workbench and up to 10 hours for meeting recordings. - Data Retention: Anonymous uploads and outputs are automatically deleted 24 hours after processing and are never used for model training. - Developer Console & API Docs: https://console.dataarctech.com/en · https://console.dataarctech.com/en/api-playground ## Key Verified Benchmarks (Tested 2026-11-11) - Public Thai ASR Suite (FLEURS, Common Voice 17, GigaSpeech 2): DataArc-ASR-TH achieves 92.79% recognition accuracy vs. ElevenLabs Scribe v2 (91.68%) and Google Chirp v3 (91.44%). - In-House 500-Sample Real-World Thai ASR Benchmark (10 domains, PyThaiNLP newmm + NFKC normalization): DataArc-ASR-TH-async achieves 3.68% P50 CER / 5.88% Macro CER / 14.68% P90 CER; DataArc-ASR-TH-streaming achieves 44.03 ms first-token latency and 0.06 P50 RTF (self-hosted GPU 2.3 GB / CPU). - Thai TTS Subjective Benchmark (100 utterances, blind 5-point evaluation by 4 native Thai speakers with independent per-annotator Z-score normalization): DataArc-TTS-TH ranks #1 with 4.41/5 weighted score and 4.39/5 MOS, ahead of ElevenLabs v3 (4.33 / 4.32) and Google Gemini TTS (4.29 / 4.27). ## Products & Online Workbench - [Real-Time Speech Recognition & TTS](https://www.dataarc.tech/en/playground/realtime-asr): Live browser microphone ASR for Thai, Chinese, English, and mixed speech, plus preset voice synthesis. - [ArcNotes · AI Meeting Notes](https://www.dataarc.tech/en/products/meeting-note): Multi-speaker transcription with speaker diarization, summaries, decisions, action items, Markdown/Word export, and click-to-play timestamps (App: https://app.arcnotes.ai/). - [ArcDub · AI Video Dubbing](https://www.dataarc.tech/en/products/dubbing): End-to-end transcription, translation, voice-cloned dubbing, bilingual subtitles, and lip sync (App: https://app.arcdub.com/). - [Multilingual Subtitles](https://www.dataarc.tech/en/products/subtitle): Interactive 8-language subtitle player and timeline fusion preview. - [Free Online Transcription & TTS Tools](https://www.dataarc.tech/en/tools): No-signup file transcription (TXT / SRT / VTT export) and Thai WAV speech synthesis. - [Developer API Playground & Docs](https://console.dataarctech.com/en/api-playground): REST, WebSocket streaming ASR, and TTS API documentation and key management. ## Benchmarks & Open Evaluation Data - [Public Benchmarks Portal](https://www.dataarc.tech/en/benchmarks): Full ASR public/in-house tables, 10-domain breakdown, code-switching CER, Z-score normalized Thai TTS evaluation, and methodology. - [Raw Benchmark Results CSV](https://www.dataarc.tech/downloads/benchmarks-results.csv): Downloadable UTF-8 CSV containing all settled public ASR, 500-sample in-house ASR, and Thai TTS Z-score benchmark datasets. ## Technical Blog - [Thai has no spaces: how speech recognition handles word segmentation and evaluation](https://www.dataarc.tech/en/blog/thai-no-spaces) (2026-10-23): Thai has no spaces between words, so WER shifts with the tokenizer. Why we use character error rate as the main metric, and how we normalize text first. - [Turning a Thai meeting into actionable minutes](https://www.dataarc.tech/en/blog/thai-meeting-minutes) (2026-05-19): From recording and speaker separation to the summary and action items: how a 40-minute Thai meeting becomes a set of minutes. - [Thai, English and Chinese in one sentence: why code-switching is hard to recognize](https://www.dataarc.tech/en/blog/zh-en-th-code-switching) (2025-12-08): Where models go wrong when Thai speech contains English brand names and Chinese personal names, and how to score the results. - [Streaming recognition and offline refinement: balancing latency and accuracy](https://www.dataarc.tech/en/blog/streaming-and-offline-refine) (2025-10-14): Streaming recognition shows text as you speak; an offline model refines each sentence once it ends. How the two results merge and how the interface transitions. - [Measuring first-token latency: how we define and test it](https://www.dataarc.tech/en/blog/measuring-first-token-latency) (2025-08-26): From the first audio frame in to the first character out: where the clock starts, how network and hardware are fixed, and what P50 and P90 each tell you. - [Voice cloning from an 8-second sample: recording tips](https://www.dataarc.tech/en/blog/voice-clone-8s-reference) (2025-06-11): A quiet room, steady volume, and a natural pace—the quality of your 8-second sample sets the ceiling for the cloned voice. - [Dubbing short videos into Thai: from subtitles to voice cloning](https://www.dataarc.tech/en/blog/thai-short-video-dubbing) (2025-04-17): Transcription, translation, voice cloning, and duration alignment: the end-to-end workflow for localizing a short e-commerce video from Chinese into Thai. ## Localized Sites & Thai SEO Guides - [English Home](https://www.dataarc.tech/en) - [Thai Home](https://www.dataarc.tech/th) · [Thai Sitemap](https://www.dataarc.tech/th/sitemap) - [Simplified Chinese Home](https://www.dataarc.tech/zh-CN) - [Thai Speech-to-Text Guide](https://www.dataarc.tech/th/speech-to-text) · [Thai ASR API Guide](https://www.dataarc.tech/th/asr-api) · [Thai TTS API Guide](https://www.dataarc.tech/th/tts-api) · [Thai Meeting Transcription Guide](https://www.dataarc.tech/th/meeting-transcription) - [Contact & Company Info](https://www.dataarc.tech/en/contact) ## Full Machine-Readable Context - [Full DataArc Context & Benchmark Tables](https://www.dataarc.tech/llms-full.txt): Complete evaluation tables, Thai solution catalog, and full text of all published engineering articles. --- ## Complete Benchmark Tables (Tested 2026-11-11) ### 1. Public & TTS Core Overview | Task | Dataset | Metric | DataArc | Competitor A | Competitor B | | --- | --- | --- | --- | --- | --- | | ASR | FLEURS · Common Voice 17 · GigaSpeech 2 | accuracy (higher is better) | DataArc-ASR-TH: 92.79% | ElevenLabs Scribe v2: 91.68% | Google Chirp v3: 91.44% | | TTS | 泰语主观评测集 | weighted-score (/5) (higher is better) | DataArc-TTS-TH: 4.41 | ElevenLabs v3: 4.33 | Google Gemini TTS: 4.29 | | TTS | 泰语主观评测集 | mos (/5) (higher is better) | DataArc-TTS-TH: 4.39 | ElevenLabs v3: 4.32 | Google Gemini TTS: 4.27 | ### 2. In-House Real-World Thai ASR (500 Samples, 10 Domains) Evaluation pipeline: `cal_wer_cer.py` (Unicode NFKC, canonical Thai vowel/tone ordering, regex normalization for numbers/currencies/dates, and `PyThaiNLP newmm` segmentation). | Dimension | Google Gemini 3.5 Transcribe | DataArc-ASR-TH-async | DataArc-ASR-TH-streaming | Cartesia ink-whisper | Notes | | --- | --- | --- | --- | --- | --- | | Median CER (P50) | 2.82% | 3.68% | 7.46% | 17.06% | Gemini 3.5 ranks first in median precision, with 50% of clips at or below 2.82% CER | | Mean CER (Macro CER) | 6.23% | 5.88% | 9.46% | 19.27% | DataArc-ASR-TH-async and Gemini sit in the top tier, well ahead of Cartesia | | P90 CER | 16.74% | 14.68% | 23.69% | 35.96% | DataArc-ASR-TH-async and Gemini remain resilient on long-tail clips with background noise | | Median WER (P50) | 5.00% | 5.26% | 11.76% | 30.00% | Gemini achieves the best word-level median accuracy (5.00%) | | Mean WER (Macro WER) | 8.03% | 7.42% | 12.50% | 31.46% | DataArc-ASR-TH-async and Gemini perform closely on macro word error rate | | Zero-error clips (0 CER) | 233 / 500 (46.6%) | 192 / 500 (38.4%) | 100 / 500 (20.0%) | 36 / 500 (7.2%) | Nearly half of Gemini's transcripts have zero character errors | | Excellent clips (CER ≤ 5%) | 300 / 500 (60.0%) | 278 / 500 (55.6%) | 185 / 500 (37.0%) | 72 / 500 (14.4%) | 60% of clips are transcribed with virtually no noticeable errors | | Good clips (CER ≤ 10%) | 386 / 500 (77.2%) | 400 / 500 (80.0%) | 330 / 500 (66.0%) | 130 / 500 (26.0%) | The vast majority of samples fall into the production-ready range | | Recognition success rate | 500 / 500 (100.0%) | 500 / 500 (100.0%) | 500 / 500 (100.0%) | 500 / 500 (100.0%) | All 500 utterances completed without timeouts or dropped streams | | Average response latency | 4.64 s (P50: 4.57s) | Full-clip upload (1-3s) | 44.03 ms (first-token) | 0.26 s (P50: 0.23s) | Gemini runs non-streaming LLM inference with higher end-to-end latency | | Single-request RTF (P50) | 0.96 | 0.18 | 0.06 | 0.05 | DataArc-ASR-TH-streaming and Cartesia deliver ultra-fast real-time factors | | Deployment & operations | Cloud API (managed) | Self-hosted GPU (13.4 GB VRAM) | Self-hosted GPU (2.3 GB) / CPU | Cloud API (managed) | Gemini and Cartesia require zero local GPU infrastructure | #### Domain Breakdown (P50 CER) | Domain | Samples | Gemini 3.5 | DataArc-ASR-TH-async | DataArc-ASR-TH-streaming | Cartesia | | --- | --- | --- | --- | --- | --- | | Technology | 56 | 2.27% | 6.28% | 11.63% | 11.51% | | Legal & Government | 4 | 2.00% | 2.67% | 3.83% | 13.53% | | Finance | 147 | 2.67% | 2.63% | 5.56% | 20.00% | | Daily Conversation | 202 | 3.80% | 4.51% | 7.58% | 16.77% | | Entertainment | 8 | 1.06% | 3.12% | 6.32% | 11.59% | | Travel & Culture | 7 | 0.00% | 3.90% | 9.09% | 22.08% | | Other (Creator & Media) | 54 | 0.00% | 4.21% | 7.50% | 16.95% | | Interview & Podcast | 4 | 5.57% | 3.61% | 5.83% | 18.29% | | Education | 6 | 8.93% | 4.57% | 8.58% | 9.19% | | News | 5 | 9.86% | 2.50% | 5.26% | 18.42% | #### Pure Thai vs. Thai-English Code-Switching | Subset | Samples | Gemini 3.5 (Mean / P50 CER) | DataArc-ASR-TH-async (Mean CER) | DataArc-ASR-TH-streaming (Mean CER) | | --- | --- | --- | --- | --- | | Pure Thai (Non Code-Switching) | 350 | 4.91% / 0.00% | 5.40% | 8.77% | | Thai-English (Code-Switching) | 150 | 9.29% / 7.00% | 7.00% | 11.08% | ### 3. Thai TTS Subjective Benchmark (100 Utterances, Z-Score Normalized) Weighted score dimensions: Prosody (25%), Text fidelity (20%), Clarity (20%), Emotional consistency (20%), Pauses & breath (10%), Speaking rate (5%). Global baseline across 4 native Thai raters: mean 3.9316, std 0.7789. | Rank | Model | Weighted (/5) | MOS (/5) | Text Fidelity | Clarity | Speaking Rate | Prosody | Emotion | Pauses & Breath | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | 1 | DataArc-TTS-TH | 4.41 | 4.39 | 4.53 | 4.72 | 4.23 | 4.12 | 4.35 | 4.33 | | 2 | ElevenLabs v3 | 4.33 | 4.32 | 4.61 | 4.54 | 4.27 | 3.88 | 4.33 | 4.30 | | 3 | Google Gemini TTS | 4.29 | 4.27 | 4.56 | 4.54 | 4.04 | 4.04 | 4.23 | 4.16 | | 4 | Google Chirp 3 HD | 4.07 | 4.03 | 4.52 | 4.61 | 3.66 | 3.77 | 3.73 | 3.85 | | 5 | CosyVoice 2 | 3.72 | 3.76 | 3.73 | 3.71 | 3.96 | 3.58 | 3.80 | 3.80 | | 6 | Thonburian | 3.72 | 3.74 | 3.00 | 3.31 | 4.04 | 4.17 | 4.17 | 3.72 | --- ## Thai Product & Scenario Landing Pages ### Product Solutions - **Thai Speech to Text** (ถอดเสียงพูดภาษาไทยเป็นข้อความ) — https://www.dataarc.tech/th/speech-to-text - ถอดเสียงภาษาไทยเป็นข้อความออนไลน์ฟรี ไม่ต้องติดตั้งโปรแกรมหรือสมัครสมาชิก ใช้ได้กับไฟล์ประชุม สายสนทนา และวิดีโอ เมื่อใช้งานจริงเชื่อมต่อผ่าน ASR API ได้ - **Thai Audio to Text** (แปลงไฟล์เสียงเป็นข้อความ) — https://www.dataarc.tech/th/audio-to-text - อัปโหลดไฟล์เสียงหรือวิดีโอภาษาไทย แล้วรับข้อความพร้อมเวลากำกับ เหมาะกับไฟล์ประชุม สัมภาษณ์ สายสนทนา และวิดีโอ ถอดเสียงจากไฟล์ผ่าน API ได้เมื่อใช้งานจริง - **Thai Text to Speech** (แปลงข้อความเป็นเสียงพูดภาษาไทย) — https://www.dataarc.tech/th/text-to-speech - พิมพ์ข้อความภาษาไทย เลือกเสียง ฟังตัวอย่าง แล้วดาวน์โหลดเป็นไฟล์เสียงได้ฟรี ประเมินคุณภาพเสียงก่อนเชื่อมต่อ TTS API ภาษาไทย - **Thai ASR API** (ASR API ภาษาไทย) — https://www.dataarc.tech/th/asr-api - เชื่อมต่อ ASR API ภาษาไทยเพื่อแปลงเสียงพูดเป็นข้อความในแอปและระบบของคุณ ทดลองผลลัพธ์บนเว็บก่อน แล้วค่อยขอ API Key - **Thai TTS API** (TTS API ภาษาไทย) — https://www.dataarc.tech/th/tts-api - ใช้ TTS API ภาษาไทยสร้างเสียงพูดจากข้อความสำหรับผู้ช่วยเสียง งานคอนเทนต์และสื่อออนไลน์ บทเรียนออนไลน์ และระบบบริการลูกค้า ทดลองเสียงบนเว็บได้ก่อน - **Real-time ASR** (ถอดเสียงแบบเรียลไทม์ (Real-time ASR)) — https://www.dataarc.tech/th/realtime-asr - ถอดเสียงภาษาไทยแบบเรียลไทม์ ข้อความขึ้นทันทีตามเสียงพูด สำหรับคำบรรยายสด ผู้ช่วยเสียง และคอลเซ็นเตอร์ ทดลองบนเว็บก่อนเชื่อมต่อ API แบบสตรีมมิง ### Industry & Workflow Scenarios - **Call Center Transcription** (ถอดเสียงสายคอลเซ็นเตอร์) — https://www.dataarc.tech/th/call-center-transcription - ถอดเสียงสายคอลเซ็นเตอร์ภาษาไทยเป็นข้อความ สำหรับตรวจคุณภาพบริการ บันทึกเคสใน CRM และสรุปบทสนทนา ทดลองถอดเสียงก่อนเชื่อมต่อ API - **Meeting Transcription** (ถอดเสียงประชุมเป็นข้อความ) — https://www.dataarc.tech/th/meeting-transcription - ถอดเสียงประชุมภาษาไทยเป็นข้อความ สำหรับทำรายงานการประชุม ติดตามงาน และค้นหาย้อนหลัง ทดลองออนไลน์ก่อนเชื่อมต่อ API - **Video Subtitles** (สร้างซับไตเติลวิดีโอภาษาไทย) — https://www.dataarc.tech/th/video-subtitles - อัปโหลดวิดีโอภาษาไทย แล้วรับไฟล์ซับไตเติล SRT ที่แบ่งช่วงเวลาตามประโยค ใช้กับสื่อออนไลน์ คอร์สเรียน และคลังวิดีโอของทีม - **Voice Assistant** (ผู้ช่วยเสียงภาษาไทย) — https://www.dataarc.tech/th/voice-assistant - สร้างผู้ช่วยเสียงภาษาไทยด้วยระบบถอดเสียงแบบเรียลไทม์และสร้างเสียงพูดโต้ตอบ รับเสียงผู้ใช้ แปลงเป็นข้อความ แล้วตอบกลับเป็นเสียงภาษาไทย - **Customer Service AI** (AI ด้านเสียงสำหรับบริการลูกค้า) — https://www.dataarc.tech/th/customer-service-ai - ใช้ ASR และ TTS ภาษาไทยยกระดับระบบบริการลูกค้าด้วย AI ถอดเสียงบทสนทนา วิเคราะห์คุณภาพบริการ และตอบกลับลูกค้าด้วยเสียงที่เป็นธรรมชาติ - **Education TTS** (เสียงอ่านบทเรียนออนไลน์) — https://www.dataarc.tech/th/education-tts - แปลงข้อความเป็นเสียงอ่านภาษาไทยสำหรับบทเรียนออนไลน์ สื่อการเรียนรู้ และผู้เรียนที่ต้องการฟังแทนการอ่าน --- ## Full Technical Blog Articles ### Thai has no spaces: how speech recognition handles word segmentation and evaluation - URL: https://www.dataarc.tech/en/blog/thai-no-spaces - Category: tech - Published: 2026-10-23 - Summary: Thai has no spaces between words, so WER shifts with the tokenizer. Why we use character error rate as the main metric, and how we normalize text first. Thai is written without spaces between words; a space usually appears only between phrases or sentences. That turns “where should the recognized text be split into words” into a separate problem, and it makes word error rate (WER), the usual metric, unstable for Thai. This article describes the metrics and the text normalization we use when we evaluate Thai recognition. > **Key points** > > - On the same Thai text, a different tokenizer can change WER noticeably. Character error rate (CER) does not depend on tokenization. > - We apply Unicode normalization before scoring, so characters that look the same but are encoded differently are not counted as errors. > - CER is our main metric. We also report WER with a fixed tokenizer, and we state the test set and version for every number. ## 1. Why Thai has no spaces Thai is a continuous script: the words in a sentence run together, and readers split them using their knowledge of the vocabulary. A space in Thai works more like a comma or a full stop. It separates phrases and sentences. For speech recognition, the model outputs one continuous string of characters. If a downstream system needs search, translation or subtitle line breaks, it has to segment the text again. Segmentation is ambiguous in itself: the same string of characters can have more than one valid split, and different tokenizers use different dictionaries and rules. - Example `ตากลม` (One string of characters, two ways to split it): - `ตา | กลม` — round eyes - `ตาก | ลม` — airing in the wind *Figure: A common example of ambiguity in Thai word segmentation.* ## 2. Why WER is unstable on Thai Word error rate counts substitutions, deletions and insertions per word, then divides by the number of words in the reference. For Thai, both the reference and the recognition result must be segmented before they can be compared, so the tokenizer becomes part of the evaluation: change the tokenizer and both the word count and the alignment change, and WER changes with them. Character error rate counts per character and needs no segmentation. The result depends only on the two texts themselves. That is why CER is the more common main metric for Thai recognition. ```text CER = (S + D + I) / N S substituted characters D deleted characters I inserted characters N characters in the reference text ``` CER has limits too. It treats a wrong character and a wrong word alike, and it cannot show directly whether a sentence reads naturally. So we also report WER under a fixed tokenizer, for reference. ## 3. Text normalization before evaluation In Thai, vowel and tone marks are written above or below the consonant. Text that looks the same can be encoded in different ways in Unicode: - The vowel ำ (U+0E33) can also be written as ํ (U+0E4D) followed by า (U+0E32). The two display almost identically. - If a tone mark and an upper vowel are typed in a different order, the code point sequence is different. - Thai digits ๑๒๓ and Arabic digits 123 are mixed, and so are full-width and half-width symbols. Unless these are unified first, every such difference is counted as a recognition error. Before computing metrics we apply the same processing to the reference and the recognition result: Unicode normalization, canonical Thai vowel and tone ordering, unified digits and punctuation, and removal of extra whitespace. ## 4. What we do 1. The main metric is CER, computed after the normalization above. 2. The reference metric is WER, always segmented with `PyThaiNLP newmm`. 3. Every number states the test set, model version, hardware and test date. 4. The full evaluation methodology and raw results CSV are available on the [public benchmarks](https://www.dataarc.tech/en/benchmarks) page. ### Turning a Thai meeting into actionable minutes - URL: https://www.dataarc.tech/en/blog/thai-meeting-minutes - Category: tutorial - Published: 2026-05-19 - Summary: From recording and speaker separation to the summary and action items: how a 40-minute Thai meeting becomes a set of minutes. A 40-minute Thai meeting recording is only raw material. Turning it into minutes that people actually read and act on takes a few human checks: how well the meeting was recorded, whether the speakers were separated correctly, which parts of the transcript deserve another listen, and whether the decisions and action items are complete. This article walks through those steps in order, with what the machine does at each one and where you should look. > **Key points** > > - Most of the quality of the minutes is decided at recording time: put the recorder in the middle of the table and ask everyone to say their name at the start. > - You do not need to read every word. Check the speakers first, then the decisions and action items, then spot-check the transcript using click-to-play timestamps. > - Every action item needs an owner and a due date. If either is missing, open its position in the recording and listen. ## 1. Recording: most of the quality is decided here The model can only work with what it hears, and the minutes can only contain what the model heard. A few things cost almost nothing and affect every later step: - Put the recording device in the middle of the table. Do not leave it in a bag or under a laptop. Voices from far away get smeared by the echo of the room. - When someone joins remotely, make sure their voice goes into the same recording, at a volume close to the people in the room. - At the start, ask everyone to say their name and role in one sentence. Later, this lets you match "Speaker A, B, C" to real people quickly. - Try not to talk over each other. When two people speak at once, both the transcript and the speaker separation tend to go wrong. - Thai meetings often include English product names and abbreviations such as KPI and OKR. Say the full name the first time and use the short form afterwards. It saves a lot of work during review. When you upload, prepare the file according to the formats and the per-meeting length limit shown on the page. ## 2. What happens after you upload The minutes are generated in four stages: uploading the recording, transcribing speech, separating speakers, and writing the summary and action items. The page shows the current stage and progress, so you do not need to watch it. Processing time depends on the length of the meeting; the figures are in the metrics section of the [meeting notes page](https://www.dataarc.tech/en/products/meeting-note). If the connection drops, your daily free limit is reached, or the recording exceeds the per-meeting duration limit, the page shows a clear explanation and a retry option. When it finishes you get four things: a transcript with speakers and timestamps, a summary, decisions and action items. We suggest reading them in reverse: check whether the action items and decisions make sense first, and look into the transcript only where you have doubts. ## 3. Check the speakers first Speakers are first shown as letters: Speaker A, B, C. Start by counting. Does the number match the people who actually attended? If four people attended and the result shows six, one attendee was split into two speaker tracks. Then listen to the first thing each person says, match it to a real person and rename the speaker. Every action item owner later uses that name. There are two common problems: - One person is split into two speakers. This often happens when they change seats halfway through, or their distance from the microphone changes. You can merge the two sets of segments. - Short replies such as "ครับ" and "ค่ะ" are easily assigned to the previous speaker. This does not affect the summary, but be careful when you quote who said what. Speaker separation is measured with DER: ```text DER = (missed speech + false alarm + confusion) / total speech time ``` Missed speech means someone spoke but was not marked. A false alarm means nobody spoke but a segment was marked. Confusion means a segment was marked but given to the wrong person. DER is weighted by duration, so one long turn assigned to the wrong speaker weighs far more than a few short replies. ## 4. Review the transcript without reading every word Reading a 40-minute meeting word by word is not worth the time. Review in this order of priority: 1. Names of people, companies and products. These are the words most likely to be wrong, and they do the most damage to the credibility of the minutes. 2. Numbers, dates and amounts. Thailand commonly uses the Buddhist Era (พ.ศ.). The Buddhist year is the Gregorian year plus 543, so 2569 is 2026. If the minutes feed a system that uses Gregorian dates, convert them consistently. 3. Mixed-language words. English brand names keep their original letters and are not forced into Thai script. Where the spelling is inconsistent, change it to one form in the minutes. 4. Negations. In Thai, a missing "ไม่" reverses the meaning. Wherever you have doubts, click the item and the player jumps to that position in the recording. Listen to those few seconds only, not the whole meeting from the start. ## 5. Summary, decisions and action items: three questions The summary should answer: what did this meeting decide, and what did it leave open? Two or three sentences are enough. Decisions should only list things that were agreed. Topics that were discussed without a conclusion do not count. Every decision carries its position in the recording, so when you are unsure, listen to the tone. In a Thai meeting, "ลองดูก่อน" (let's try it first) and "ตกลงตามนี้" (agreed, let's go with this) can both sound like a commitment, but they mean very different things once they are written into the minutes. An action item must answer three questions: what is to be done, who does it, and by when. - Task: send the load-test report to the customer this week - Owner: Speaker B - Due: next Wednesday - Recording position: 00:12:40 *Figure: The fields of one action item in the minutes.* The owner is the speaker, and shows as a name once you have renamed the speaker. The due date is recorded in the speaker's own words, for example "next Wednesday". Convert it to a specific date using the meeting date, otherwise nobody will remember which day it was a week later. For an action item that lacks an owner or a due date, listen to its position in the recording. Most of the time it was said in the meeting but not recognized as part of the action item. ## 6. Export and check against the recording Export after you have reviewed. There are two formats: - Markdown: the title, duration and number of speakers, the summary, numbered decisions, action items with checkboxes, and the full transcript with speakers and times. It suits a knowledge base or a code repository. - Word: the same content, for colleagues who do not use those tools. The header of an exported file contains only the title, the duration and the number of speakers. It has no meeting date. Put the date in the title, or in the body of the email you send the minutes with, so that "next Wednesday" in an action item has something to refer to. Do one last thing before you send it: click the recording position of every decision and action item. There are not many of them and it takes a few minutes, which is much less work than being asked "who said this?" afterwards. ### Thai, English and Chinese in one sentence: why code-switching is hard to recognize - URL: https://www.dataarc.tech/en/blog/zh-en-th-code-switching - Category: tech - Published: 2025-12-08 - Summary: Where models go wrong when Thai speech contains English brand names and Chinese personal names, and how to score the results. In meetings, live streams and customer service calls in Thailand, it is entirely ordinary for a sentence to contain an English brand name, an abbreviation or a Chinese personal name. For the speaker this is no obstacle at all. For speech recognition it is a separate kind of problem. This article explains where errors in mixed-language speech usually appear, and how to score them so that a different way of writing a word is not counted as a mishearing. > **Key points** > > - Errors in mixed speech are not spread evenly. They concentrate at three places: where the language switches, on proper nouns, and in the question of which script a word should be written in. > - Scoring needs one unit first: characters for Thai and Chinese, words for English, then combined into a mixed error rate (MER). Report per-language numbers as well, so errors on proper nouns are not diluted by the large Thai denominator. > - The normalization rules must be written before the evaluation: which script loanwords use, which spellings the alias list accepts, how to treat Traditional and Simplified Chinese and full-width and half-width forms. They apply to every system equally. ## 1. What mixed speech is and why it is hard Code-switching means one speaker alternating between several languages in the same stretch of speech. When it happens inside a sentence, such as an English product name embedded in a Thai sentence, it is called intra-sentential switching, and it is the hardest kind. The difficulty comes from three sides: - Data. Most training audio has one language per recording. The model rarely sees the moment a switch happens, so it does not know what can follow a Thai word. - Pronunciation. Few consonants can end a syllable in Thai, so English words are re-pronounced to fit Thai rules once they are borrowed. For example, ball is written บอล and the final letter is read as an n sound. English words spoken by Thai speakers do not sound like recordings of native speakers, so what the model learned from English data cannot be copied over directly. - Script. The model does not only have to hear which word it is, it also has to decide which script to write it in. The same word can be written in Latin letters or in Thai, both are in use, and the reference answer will contain only one of them. ## 2. Where errors usually appear ### Brand and product names Words such as Shopee, LINE and Grab are written in Latin letters in product interfaces, and often in Thai in casual chat, for example ช้อปปี้ and ไลน์. The model may output either form, or a string of Thai words that sound close but have nothing to do with the brand. It is worse when the brand is also an ordinary word. ### Spelling of loanwords Some loanwords have a settled Thai spelling. Almost nobody writes คอมพิวเตอร์ as computer. Others have no settled form: "meeting" is written as meeting by some people and as มีตติ้ง by others. The second kind is where most disagreements in evaluation come from. ### Chinese personal names and place names When a Chinese name sits in a Thai sentence, the speaker may read it in Mandarin or in a dialect such as Teochew. Either way, the sound does not map one to one onto the characters: 李明 and 李铭 are pronounced exactly the same, and the audio alone cannot tell them apart. That takes context or a name list. ### Language boundaries Errors usually appear in the one or two words before and after a switch. When the switched part is longer, the model may also get the language wrong for the whole stretch and write a short English phrase as Thai words that sound similar. ### Numbers, abbreviations and units 5G, KPI, "สองล้าน" (two million) and "2 ล้าน" are spoken and written differently, and Thai digits ๑๒๓ are mixed with Arabic digits 123. The more languages involved, the more ways there are to write the same thing. ### Tones Thai and Chinese both have tones, but the tone systems differ. A Chinese word inserted into a Thai sentence is sometimes spoken with Thai intonation, which weakens the tone cue. The model then leans more on context, and context is thinnest exactly at the switch. ## 3. Two common designs and where each fails There are two common ways to handle multilingual speech. One is to identify the language first and then hand the audio to a model for that language. Switching between sentences is no problem for it: this sentence is Thai, the next is English, and each goes its own way. Switching inside a sentence does not work, because language identification works on a whole sentence or even a longer stretch. An English word inside a sentence is recognized as Thai and written as Thai words that sound similar. The other is one model that supports several languages, with a vocabulary that holds several scripts. It can switch inside a sentence, but words from different languages compete in the same vocabulary. When the training data has few examples with switching, the model tends to stay in the base language of the sentence and pull the embedded word toward it. Both designs need to be checked with a test set that contains switching. A good score on a single-language test set does not imply a good score on mixed speech. ## 4. Scoring: WER and CER are not enough Thai has no spaces, so WER depends on the tokenizer, as we explained in [another post on Thai word segmentation and CER](https://www.dataarc.tech/en/blog/thai-no-spaces). The same is true of Chinese. English is normally counted by word, and counting by character exaggerates the differences. In a line of mixed text no single unit suits all three scripts. The standard approach is mixed error rate (MER): segment each script using its natural unit into a single sequence of tokens, then compute edit distance across the whole sequence. ```text MER = (S + D + I) / N ``` S, D and I are the numbers of substituted, deleted and inserted tokens, and N is the number of tokens in the reference. Tokens are cut like this: - Thai: by character, with each vowel and tone mark counted as one. - Chinese: by character. - English letters and digits: a continuous run counts as one word. Here is an example of this convention. - Reference: ผมคุยกับ 李明 เรื่อง Shopee - Hypothesis A: ผมคุยกับ 李名 เรื่อง shopee. After normalization only one token differs, 名 against 明, so 1 error. Case does not count. - Hypothesis B: ผมคุยกับ 李明 เรื่อง ช้อปปี้. The sound is completely right, but the brand is written in Thai, so 7 errors: 1 substitution and 6 insertions. - Meaning: I talked with Li Ming about Shopee. *Figure: A Thai sentence with a Chinese personal name and an English brand name, N = 17: 14 Thai characters, 2 Chinese characters and 1 English word. The text is an illustration.* Hypothesis B shows why the script convention has to be fixed first. An English word takes only 1 token in the reference, so misspelling a letter costs 1. Writing it in Thai instead takes 7 characters and costs 7, even though what was heard is the same. MER also has a blind spot. Its denominator is mostly Thai characters, and English words and Chinese names make up a small share. The overall MER can look low while proper nouns are wrong in many places. So besides the overall MER, report the error rate per language, and an error rate computed only on tokens at the points where the language switches. ## 5. Normalization before scoring Normalization applies the same treatment to the reference and the hypothesis, and the rules are written before the evaluation: - Unicode is normalized with NFKC, full-width letters become half-width, and Thai diacritic ordering follows canonical Thai vowel and tone ordering. - English is lowercased. Punctuation is stripped for core lexical scoring (after normalizing full-width and half-width forms when punctuation is evaluated separately). - Chinese characters are unified to a single script variant (Simplified or Traditional). - Number forms are unified: Thai digits and Arabic digits count as the same, and for spoken and written numbers, one form is chosen in the annotation guideline. The most important part is the rule for loanwords, which we suggest writing as three lines: 1. Brand names, product names and abbreviations keep their original Latin letters. 2. Loanwords with an established Thai spelling are written in Thai, for example คอมพิวเตอร์. 3. Chinese personal names and place names are written in Chinese characters. The alias list contains only the spellings the guideline explicitly allows. It has to be fixed before the evaluation is run, used unchanged for every system, and published with the results. The alias list is the easiest place to tune a score quietly, and aliases added or removed afterwards do not count as normalization. ## 6. Building the test set - Use real conversations rather than reading from a word list. Brand names read aloud are pronounced too carefully and are far from real use. - Annotate against one guideline, with two people annotating each recording. Resolve disagreements by revising the guideline, not case by case. - For every sentence, record which languages are used, how many times it switches, and the type of switch: one word embedded, a phrase embedded, or a whole sentence. Report by these groups, so that you can see whether errors rise with more switching and which kind of switch is hardest. - Brand names and personal names in the test set must not enter the training data, and must not enter any word list used for tuning. The size, version and test date of our in-house mixed-speech set are on the [public benchmarks](https://www.dataarc.tech/en/benchmarks) page. ## 7. What we do In the product, English brand names and Chinese personal names inside a sentence are output in the language the speaker actually used, and are not forced into Thai transliteration. For evaluation, the main metric on the mixed-speech set is MER. The test set, the normalization rules and the calculation method are on the [public benchmarks](https://www.dataarc.tech/en/benchmarks) page. ### Streaming recognition and offline refinement: balancing latency and accuracy - URL: https://www.dataarc.tech/en/blog/streaming-and-offline-refine - Category: tech - Published: 2025-10-14 - Summary: Streaming recognition shows text as you speak; an offline model refines each sentence once it ends. How the two results merge and how the interface transitions. Live captions pull in two directions. The text has to appear as soon as the words are spoken, and once a sentence is finished it should be as accurate as possible. Streaming models are good at the first, offline models at the second. We use both: streaming recognition produces text while you speak, and at the end of each sentence the audio of that sentence goes to an offline model for refinement, and the refined result replaces the original text. This article covers how the two results merge, how the interface transitions between them, and where things can go wrong and how we fall back. > **Key points** > > - A streaming model can only judge from the audio it has heard so far, while an offline model sees the whole sentence. After refinement, numbers, proper nouns and punctuation are usually more reliable. > - The merge rules are simple: a partial replaces the whole current line, a final is added as a sentence, and a refinement replaces the sentence with the same number. If refinement returns nothing, the streaming text stays. > - Blue means "may still change" and the main text color means "will not change again". The fade from blue to the main color is the signal that a sentence is settled. ## 1. Why latency and accuracy pull against each other A streaming model has to produce text before the sentence is finished, so it can only rely on what it has already heard plus a very short look ahead. But many words are only settled by what follows: homophones depend on the next words, numbers depend on the unit, and since Thai has no spaces, even where a word ends depends on what comes after. An offline model reads the whole sentence at once, and most of these ambiguities disappear with the extra context. So "faster" and "more accurate" come down to the same choice: how long to wait before committing. Wait less and text appears sooner, with less to base the decision on. Wait longer and there is more evidence, but the user also stares at a blank screen for longer. Streaming recognition chooses not to wait: it gives a provisional result first, and allows it to be changed later. Latency also has to be looked at in two parts. First-token latency is the time from the start of speech to the first character on screen. Settling latency is the time from the end of a sentence until it becomes text that will not change. Refinement only affects the second one and never slows the first token. ## 2. Three kinds of text: partial, final and refined - Partial: everything heard so far in the current sentence. Each new partial replaces the previous one as a whole, it is not appended, and it can be empty. Characters already on screen can be changed. - Final: after the speaker pauses for about two seconds, the server settles the sentence. The connection stays open and waits for the next one. - Refined result: the browser keeps the audio it has just sent. When a final arrives, it cuts out that sentence, turns it into a WAV file and sends it to the offline recognition model, the same one used for file transcription, which returns text with punctuation. How "that sentence" is cut out is an implementation detail. We do not ask the server for the audio. Locally we look for the first silence of 1.5 seconds or longer, and cut from the first frame with sound up to that point, with a little margin on both sides. The rest of the audio is kept for the next sentence. - partial: ผมจะส่งรายงานให้ - final: ผมจะส่งรายงานให้ภายในวันสุก - refined: ผมจะส่งรายงานให้ภายในวันศุกร์ - meaning: I will send you the report by Friday. *Figure: Stages of one sentence. The text is an illustration of the mechanism.* In this example "ศุกร์" (Friday) and "สุก" (ripe) sound the same. While the streaming model has no more context after the word, either spelling can come out. With the whole sentence available, the offline model finds it easier to choose the right one. ## 3. Merge rules 1. A partial replaces the whole current line. It is never accumulated. 2. Each final becomes a numbered sentence, added after the existing sentences. 3. A refinement replaces by sentence number, not by position. When it returns, later sentences may already be on screen, and several refinements may be in flight at once. Each result changes only its own sentence, so it does not matter which one comes back first. 4. If refinement has no result, the streaming text stays. No request is sent when the audio is shorter than 0.25 seconds. If the response is empty, the request fails, the daily limit is reached, or nothing has come back after 12 seconds, the streaming text also stays. Refinement is an improvement, not a dependency. 5. For screen readers, a settled sentence is read out once, in a separate live region, when it arrives. Later replacements by refinement are not read again, so nobody hears the same sentence twice. ## 4. The interface transition: from blue to the main color Three states, three looks: - The sentence being spoken is blue and ends with a blinking caret. - A sentence that is final but whose refinement has not returned is still blue. - After the refinement comes back, or is given up on, the text is replaced and the color fades to the main color over about 0.8 seconds. Color expresses how certain the text is, not what kind of sentence it is. When readers see blue, they know it may change. When they see the main color, they know it will not. If text is swapped quietly with no signal, readers think they misread it. Only the color moves, not the layout. The caption area shows just the latest few lines, so older sentences scroll out of view naturally, and the text that gets changed is always the newest one or two sentences, which is where the reader is already looking. When the system is set to reduce motion, the caret stops blinking. ## 5. Where it goes wrong - Client and server disagree on sentence boundaries. The server settles a sentence using its own silence detection, while the client slices audio at 1.5 seconds of silence. If a sentence has a longer pause in the middle, or the server cuts in the middle of continuous speech, the audio sent for refinement may include a bit too much or too little. A safer approach is to align using the sentence start time returned by the server. That requires first confirming that the two clocks still agree after a reconnection, so for now we cut by pauses. - Long continuous speech. Only the most recent 30 seconds of audio are kept locally. For a longer sentence, the beginning does not go into refinement, and the refined text may miss the start of the sentence. - Very quiet speech. Frames below the noise threshold do not count as sound, so the whole sentence may be treated as silence. There is no audio to refine, and the streaming text stays. - Interrupted connection. When the server closes the connection normally, the web page reconnects automatically, up to three times in a row. When the connection drops abnormally, it does not reconnect and shows a network error straight away. - Trial time limit. When it is reached, the server signals it with a dedicated close code, and the page shows a specific explanation instead of a vague "something went wrong". ## 6. How to measure it First-token latency is counted like this: ```text first-token latency = time the first non-empty partial arrives − start of the first frame with sound ``` The timing is done in the browser, so it covers three parts: capture buffering, the network and server processing. Capture uses frames of 2048 samples, so a frame's length depends on the browser's sample rate: 128 milliseconds at 16 kHz, about 43 milliseconds at 48 kHz. Refinement time is also a browser-side round trip: encoding the WAV, uploading, recognizing and returning. The processing time reported by the server itself is shorter. To measure what refinement gains, score the two results separately: the final text as it was settled, and the text after refinement, each with CER on the same test set. The difference between them is the improvement refinement brings. Looking only at the accuracy of the final captions cannot show this. The numbers, test sets and hardware are on the [public benchmarks](https://www.dataarc.tech/en/benchmarks) page. ### Measuring first-token latency: how we define and test it - URL: https://www.dataarc.tech/en/blog/measuring-first-token-latency - Category: benchmark - Published: 2025-08-26 - Summary: From the first audio frame in to the first character out: where the clock starts, how network and hardware are fixed, and what P50 and P90 each tell you. First-token latency in streaming recognition is often summed up in one line: how long after you start speaking the first character appears. That leaves three things open: where the clock starts, where it stops, and under what conditions the test ran. Change the conditions and the same system can produce very different numbers. This article explains how we measure first-token latency in our benchmark table, and how it differs from the live reading on the [real-time speech page](https://www.dataarc.tech/en/playground/realtime-asr). > **Key points** > > - In the benchmark table, first-token latency runs from the moment the first audio frame is sent to the recognition service until the first recognized character comes back. > - The test runs inside one data center. GPU, concurrency, model version, test date and test set are fixed and published with the number. We report P50 and P90 together, not just one of them. > - The "First token" reading on the real-time speech page is measured in the browser. It includes capture buffering and your own network, so it cannot be compared directly with the table. ## 1. How we define first-token latency A streaming recognition server returns partial results while it is still receiving audio. A partial result is the text recognized so far for the sentence being spoken, and it may be revised later. First-token latency looks only at the first partial result whose text is not empty, that is, the first recognized character. An empty result does not count, because the user still sees nothing. ```text first-token latency = time the first character returns − time the first frame is sent ``` ## 2. Why we measure inside the data center The clock does not start when someone taps the microphone. The user may not have started speaking yet, and counting from the tap would measure human reaction time. So no person takes part in the test: a test program sends the audio of the test set to the service frame by frame. The test program and the recognition service run in the same data center, so the network round trip is short. What the number mainly measures is the server's time from receiving audio to returning the first character: decoding, recognition and producing the first result. Capture buffering on the user's device and the network between the user and the data center are not in this number. ## 3. Which conditions are fixed A different GPU, or a few more concurrent streams, changes first-token latency. We fix the following conditions during evaluation and publish them with the number: - Server GPU model and memory configuration - Number of concurrent streams - Model version and test date - Test set name and sample count If any one of these is missing, the number cannot be compared directly with numbers from elsewhere. ## 4. What P50 and P90 each tell you We measure one first-token latency for every test clip and sort all results from smallest to largest. ```text P50: 50% of runs are this fast or faster P90: 90% of runs are this fast or faster ``` P50 is the median and reflects typical performance. P90 captures tail latency: one run in ten is slower than this mark. Users notice the slow runs, so looking only at the average or P50 makes a system look smoother than it feels in practice. Averages get skewed upward by occasional outliers while masking tail latency behind fast runs, so we do not use them. The latency table on the [public benchmarks](https://www.dataarc.tech/en/benchmarks) page reports both P50 and P90. ## 5. The reading on the real-time speech page When you speak on the real-time speech page, the "First token" chip shows the value measured in your session. Before you start speaking, it shows the P50 from the benchmark table. That value is timed in the browser with `performance.now()`, and it is measured differently from the table: - The clock starts at the first audio frame that carries speech energy. When the normalized root mean square (RMS) of a frame reaches 0.015, someone is speaking, and the beginning of that frame is the start. If no frame in the whole session reaches the threshold, the start falls back to the beginning of the first frame sent. - Capture buffering is included. The browser collects 2048 samples before it hands over a frame: 128 ms at 16 kHz, about 43 ms at 48 kHz. The start is the beginning of the frame, so a whole frame has already passed by the time it is delivered. - The network between you and the server is included, and so is server processing. So the reading reflects your device and network at that moment. It is useful for getting a feel for the whole path, not for comparing with the table. If background noise is loud enough to cross the threshold, the start moves earlier, to where the noise begins, and the reading comes out higher. ## 6. Questions to ask about anyone's numbers - When does the clock start: at the tap, at the first audio frame, or at the first frame that contains speech? - When does it stop: at the first character, or at the final result for the sentence? - Where was it measured: inside the data center, or in the user's browser? On what hardware? - Is it P50, P90 or a mean? When all four questions have answers, two numbers can be compared. To see the reading on your own device, speak a few sentences into the microphone on the [real-time speech page](https://www.dataarc.tech/en/playground/realtime-asr). ### Voice cloning from an 8-second sample: recording tips - URL: https://www.dataarc.tech/en/blog/voice-clone-8s-reference - Category: tutorial - Published: 2025-06-11 - Summary: A quiet room, steady volume, and a natural pace—the quality of your 8-second sample sets the ceiling for the cloned voice. Voice cloning needs a reference recording. The model takes more from it than the timbre: it also picks up the way you speak and the room you were in. So the quality of the reference decides the ceiling for the cloned voice, and no amount of tuning afterwards can recover what the recording lacks. This article explains how to record those 8 seconds. > **Key points** > > - The reference is an 8-second mono WAV. A quiet room, steady volume and a natural pace matter more than a more expensive microphone. > - We get the reference text by running speech recognition on your recording. If you mumble, the text will be wrong, and the synthesis will follow it. > - Clone only your own voice. Before recording, the page asks you to confirm that the voice is yours and that you agree to its use for this synthesis. ## 1. How 8 seconds become a voice The process has three steps: 1. In the recording window, tick the confirmation and start recording. The browser records 8 seconds and saves them directly as a 24 kHz, 16-bit, mono WAV. No compressed format is involved, so every browser produces the same kind of file. 2. The recording is uploaded to the server and transcribed by speech recognition. The result is the reference text. If no speech can be recognized, the recording is rejected and you record again. 3. For synthesis, the recording and its reference text go to the model together with the text to be read. The output is a 24 kHz WAV, with at most 2500 characters per request. The model is not trained for you separately. The voice comes from this recording and its text, so both have to be accurate. ## 2. The recording environment - Quiet: close doors and windows, and switch off air conditioning and fans. Bare rooms with hard walls have a lot of reverb. A place with curtains or clothes is better. - Distance: keep your mouth about a fist from the microphone, and keep it the same for the whole take. Do not walk while you speak. - Volume: speak at your usual volume. Do not swing between loud and soft, and do not get so loud that it clips. Once clipping is in the recording, every synthesized sentence carries it. - Device: a headset microphone or a phone microphone is fine. Turn off voice changers and vocal enhancement filters. ## 3. How to speak The prompt in the recording dialog simply asks you to speak naturally for 8 seconds at your normal pace—say whatever you like. A few extra habits make the result more consistent: - Say complete sentences, not a list of words. Let the sentences rise and fall naturally. - Do not perform. The tone of the reference usually carries over into every synthesis. - Do not pause for long in the middle, and do not wait until the last second to begin. - Re-record if you cough, clear your throat, laugh, or other people speak in the background. ## 4. Listen first, then decide After recording, listen back once. Make sure every word is clear, with no audible background hiss, clipping, or other voices, then click "Use this clip". If anything sounds off, re-record. A fresh take takes only seconds—far easier than trying to fix synthesized audio later. ## 5. Common problems - The page says it could not hear speech: usually you are too far from the microphone, or the room is too noisy. Move closer, or go somewhere quiet, and record again. - The synthesized voice sounds muffled or echoey: most likely the reference has reverb, or you were too close to a wall. Try another room. - There is noise in the synthesis: listen back to the reference. The noise is probably already in it. Re-recording is more direct than post-processing. - The synthesized tone is more dramatic than your own: the tone in the reference was learned along with the voice. Record again in your normal tone. ## 6. A voice is personal A voice can identify a person, so we accept only the speaker's own recording, and we ask you to confirm before recording. Do not use someone else's voice, and do not upload a recording you are not authorized to use. ### Dubbing short videos into Thai: from subtitles to voice cloning - URL: https://www.dataarc.tech/en/blog/thai-short-video-dubbing - Category: tutorial - Published: 2025-04-17 - Summary: Transcription, translation, voice cloning, and duration alignment: the end-to-end workflow for localizing a short e-commerce video from Chinese into Thai. Turning a Chinese product video into a Thai version takes more than translation. The translated lines have to finish in the time the original took, the voice should still be the same person's, the subtitles have to fit, and the lip movement should not be obviously off. This article follows the processing order and covers the stages a video goes through from Chinese to Thai, and what tends to go wrong in each. > **Key points** > > - Processing has six stages: upload, speech recognition, translation, voice-cloned dubbing, lip sync, and mix and export. Errors in early stages carry through to the final video, so proofread before moving on. > - Thai and Chinese pack information differently, and the translation is often longer. Control duration at the translation stage instead of speeding up the dub later. > - The voice comes from each speaker's own voice in the original video. The cleaner the reference segment, the closer the result. ## 1. The six stages A video goes through these stages, from upload to the finished file: 1. Upload: the video is sent to the server. 2. Speech recognition: sentence-level timeline and speaker separation. 3. Translation: keep the terminology, control the length of the translation. 4. Voice-cloned dubbing: each speaker keeps their original voice. 5. Lip sync: sync lip movements to the new audio. 6. Mix and export: mixing, burned-in subtitles and the finished video. Later stages build on the results of earlier ones. If a problem shows up, go back to the stage that caused it and fix it there. Do not patch it in a later stage. ## 2. Prepare the original first The state of the original video already decides half of the result: - The voices must be clear. Recognition suffers when background music covers speech. If there is a version of the voice-over without music, use it. - Only one person should speak at any moment. People talking over each other make speaker separation and voice cloning unstable. - For lip sync, keep the face towards the camera and the mouth unobstructed. Before processing the whole video, pick a 10-to-15-second clip with two speakers and run it through every stage. Problems in recognition, translation, dubbing and subtitles show up in minutes on a short clip. Fix them, then process the full video. ## 3. Recognition: a sentence-level timeline The first step is speech recognition. It outputs the text of each sentence, a timeline, and who is speaking. The timeline is the frame for everything that follows: each translated sentence has to go back into the time window of the sentence it came from. The sentence boundaries that recognition gives are not always the right subtitle boundaries. Split sentences that are too long and merge fragments that are too short. Translation works sentence by sentence, so how cleanly the sentences are cut affects how well the translation reads. Brand names, model numbers and figures are what recognition gets wrong most often in product videos. Proofread them one by one in the subtitles before dubbing. Changing one character at this stage costs far less than changing one in the finished video. ## 4. Translation: consistent terms, controlled length The translation stage does two things: it keeps terminology, and it controls the length of the translation. For terminology, brand names, product names and model numbers should either stay as they are or have one agreed rendering, the same throughout the video. Length needs more care. Each Chinese character is one syllable and carries a lot of meaning, so the same content in Thai usually needs more syllables and takes longer to say. If the translation is too long, the dub can only be sped up, and once the speech is fast it stops sounding natural. A product video also has content that should be decided in advance: prices, currencies, units and promotional phrases. Whether to change them to what Thai viewers are used to is not something translation can decide for you. Write the decision into the glossary so the translation stays consistent. So when the translation is too long, rewrite it first: remove repetition from spoken phrasing, split a long sentence into two short ones, use a shorter way of saying it. Whatever can be solved in translation should not be left for the dubbing stage. ## 5. Voice-cloned dubbing: each speaker keeps their own voice Voice-cloned dubbing takes a reference voice from the original video for each speaker, and reads the Thai translation in that voice. The requirements for the reference segment are the same as for standalone voice cloning: one speaker, quiet room, no background music, natural pace. If a speaker only appears in segments covered by music, the cloned voice will suffer. For recording tips, see [Voice cloning from an 8-second sample: recording tips](https://www.dataarc.tech/en/blog/voice-clone-8s-reference). Dubbing naturalness is measured by Mean Opinion Score (MOS) on a 5-point cross-lingual scale. See the metrics strip on the [video dubbing page](https://www.dataarc.tech/en/products/dubbing). ## 6. Duration alignment Every Thai line has to finish within the time window of the original sentence. Otherwise the sound and the picture drift apart. To align one line of dubbing, first work out the ratio it needs: ```text ratio = natural length of the Thai line ÷ time window of the source line ``` A ratio of 1 or less means it fits, and a short pause remains after the line. A ratio slightly above 1 can be handled by adjusting the speed within a natural range, or by using the gap between two neighboring sentences. A ratio clearly above 1 means going back to translation and rewriting that sentence. Do not speed up the whole track just to make the time fit. Speeding up everything makes every sentence sound rushed, while rewriting touches only the sentence that has the problem. ## 7. Lip sync Lip sync matches lip movements naturally to the new dubbed audio. It works best for talking-head and interview videos. If the speaker's mouth is not the main thing on screen, as in product close-ups and shots of hands at work, lip sync gains little. For a video made mostly of such shots, you can skip lip sync and only replace the audio track. After it runs, check two things: at the moment speech starts and ends, whether the mouth opens early or stays open too long; and at bilabial sounds such as b, p and m, whether the lips close. Shots of a face in profile, with the mouth covered, or with a very small mouth in frame gain little, and you can skip them. ## 8. Subtitles and export The last stage mixes the tracks, burns in the subtitles and outputs the finished video. You can also download only the SRT subtitle file and use it in your own post-production. When mixing, the Thai dub should be about as loud as the voice in the original, and the music and ambient sound should stay where they were, not covering the voice. Each subtitle has to stay on screen long enough for viewers to read it. For the same meaning, a Thai subtitle has far more characters than a Chinese one, so Thai subtitles often have to be split into more, shorter pieces. Thai subtitles also differ from Chinese in two other ways: - Thai is written without spaces. Line breaks belong between words, never inside a word. Keep each line short. - Thai tone marks and upper vowels stack above the consonant, and tight line spacing cuts them off. When you burn in subtitles, use a font that supports Thai and leave a little more line spacing than for Chinese. ## 9. Check before delivery - Are brand names, model numbers and figures read correctly? - Does every line finish within the time window of the original sentence? - Are the voices of different speakers kept apart? - Are there typos in the subtitles, and do line breaks fall at word boundaries? - Is the dub balanced with the background music? - Has a native Thai speaker listened to the whole thing? Correct grammar does not mean it sounds natural.