Chinda ASR Pro v2 for recordings, Chinda ASR Realtime for live speech
Chinda ASR now comes as two products. Chinda ASR Pro v2 transcribes recordings and labels every speaker; Chinda ASR Realtime writes Thai speech down while people are still talking. In round 3 of Korakot Chaovavanich's comparison of 25 Thai speech-to-text models, published on 9 November 2026, Pro v2 placed third, in a statistical tie with the first. Pro v2 is free until 30 November 2026, and Realtime is free during the launch.
Two products
A recording? Use Pro v2. Need the text while people are talking? Use Realtime.
| Chinda ASR Pro v2 | Chinda ASR Realtime | |
|---|---|---|
| For | Recordings: meetings, calls, interviews, video | Live speech: captions, call centres, voice input |
| How | One HTTPS request per file | A WebSocket stream |
| Languages | Thai, English and Chinese | Thai |
| Speakers | Labelled on every segment | Not labelled |
| Length | Up to 8 hours and 250 MB a file | Up to 4 hours a session |
| A list of names to favour | Yes, vocabulary | Not yet |
| TVSpeech, 570 clips, characters wrong | 7.28% | 12.95% |
| Price | Free until 30 November 2026, then 0.003 IC a second | Free during the launch |
Pro v2 in round 3 of Korakot Chaovavanich's Thai ASR comparison
Korakot Chaovavanich tested 25 models on real Thai speech from YouTube (news, vlogs and interviews) in the GigaSpeech 2 set, transcribed by people, with the same clips for every model. The score is the share of characters wrong, lower is better, counted after both sides are written the same way. He published a chart of the leading models and a full table of all 25; the ranks here are from the full table.
Pro v2 placed third, with 4.6% of characters wrong on 300 everyday clips. The first seven models are within the 95% confidence interval of the first, so he counts them as a statistical tie, and their order inside the group is not significant. On his 300 hard clips Pro v2 scores 22%, the same as second place. On price, four of the other six in the tie cost less per hour of audio, one of them at a promotional price; Pro v2 costs 20% less than second place and about a tenth of the first.
Two of his footnotes apply to Pro v2, and we repeat them here. It was trained with the GigaSpeech 2 training split, which gives it an advantage on this test set (†). iApp also had the list of test clips before the test, and told him it would use the list only for a final check (‡). Before the spelling is made the same, Pro v2 has 6.0% of characters wrong and second place 5.4%; the difference is in writing conventions such as อ่ะ, มั้ย, เค้า and ๆ, not in what was heard. His test used Pro v2 as it was before the evening of 9 November 2026, when it started writing these words in standard spelling (อะ, ไหม, เขา); on 7,672 GigaSpeech 2 dev clips that lowered its error before normalising from 5.94% to 5.53%, and left the error after normalising unchanged.
Everyday and Hard are the characters wrong on his 300 everyday and 300 hard clips, after normalising; Raw is the everyday clips before normalising; USD an hour is the price per hour of audio.
Statistical tie with first place
| # | Model | Everyday | Hard | Raw | USD an hour | Timestamps |
|---|---|---|---|---|---|---|
| 1 | Gemini 3.1 Pro (Google) | 4.4% | 21% | 6.0% | $3.82 | None |
| 2 | Paxa STT Lite (Paxa Labs) | 4.6% | 22% | 5.4% | $0.50 | Word |
| 3 | iApp Chinda ASR Pro v2 † ‡ | 4.6% | 22% | 6.0% | $0.40 | Segment |
| 4 | MAI-Transcribe-2 (Microsoft) | 4.7% | 21% | 7.0% | $0.11 | Word |
| 5 | Qwen 3.8 Omni Flash (Alibaba) | 5.0% | 23% | 6.1% | $0.12 | None |
| 6 | Scribe v2 (ElevenLabs) | 5.0% | 21% | 6.8% | $0.22 | Word |
| 7 | Gemini 3.8 Flash (Google) | 5.1% | 32%* | 6.5% | $0.32 | None |
Ranks 8 to 11
| # | Model | Everyday | Hard | Raw | USD an hour | Timestamps |
|---|---|---|---|---|---|---|
| 8 | MAI-Transcribe-1.5 (Microsoft) | 5.3% | 25% | 7.5% | $0.41 | None |
| 9 | Chirp 3 (Google) | 5.3% | 24% | 7.6% | $1.09 | Word |
| 10 | Gemini 3.5 Transcribe (Google) | 6.0% | 24% | 8.2% | $0.38 | Word |
| 11 | Soniox STT v6 (Soniox) | 6.6% | 27% | 8.6% | $0.10 | Word |
Other models from his chart
| # | Model | Everyday | Hard | Raw | USD an hour | Timestamps |
|---|---|---|---|---|---|---|
| 12 | gpt-transcribe (OpenAI) | 6.8% | 27%* | 8.8% | $0.27 | None |
| 13 | Typhoon Whisper Turbo (SCB 10X, open) † | 7.0% | – | 8.0% | Free | Segment |
| 14 | Qwen3-ASR-1.7B (Alibaba, open) | 7.0% | 28% | 8.5% | $0.03 | None |
| 17 | Gemini 3.5 Flash-Lite (Google) | 7.9% | 32% | 10.0% | $0.08 | None |
| 24 | Whisper large-v3 (OpenAI, open) | 14.9% | 38% | 17.2% | $0.05 | Segment |
| 25 | MaprangSTT (PyThaiNLP, open) | 29.8% | 66% | 29.9% | Free | Word |
- Normalising: ๆ written out as the repeated word, numbers as words, อ่ะ as อะ, มั้ย as ไหม, เค้า as เขา, นึง as หนึ่ง, งี้ as นี้, เนี่ย as นี่, ฮะ as ครับ and so on, and the fillers เอ่อ, อ่า, อืม and อะ removed, on both the reference and the output.
- Hard clips: clips most systems get wrong or disagree on, such as short clips, laughter, overlapping speech, slang and names.
- † Trained with the GigaSpeech 2 training split, an advantage on this test set. ‡ Had the list of test clips before the test. * Used to select the hard clips, so its hard-clip figure is higher than it would otherwise be.
- Prices: what is paid through OpenRouter or the provider's own API; MAI-Transcribe-2's is a promotion to the end of the year. Free means an open model you run yourself. Pro v2's $0.40 is its price from 1 December 2026, about 13.5 THB an hour billed by the second.
- Timestamps: as each provider documents them: per word, per segment (a sentence or a speaker's turn), or none.
- Eight more models in his full table are not shown here: ranks 15, 16 and 18 to 23. His summary chart leaves out MAI-Transcribe-1.5, because the same provider has a better, cheaper model with the same features.
Source: Korakot Chaovavanich, "ASR ภาษาไทย ตัวไหนแม่นสุด? (พ.ย. 2026) · รอบ 3" ("Which Thai ASR is the most accurate? November 2026, round 3"), published on his Facebook page on 9 November 2026. The tables reproduce his chart, his full table and their notes. Chinda ASR Realtime was not part of this test.
Chinda ASR Pro v2: recordings
Chinda ASR Pro v2 turns a recording into text, one timestamped segment per speaker turn.
- Accuracy: 97.94% of characters right on the 3,539 Common Voice 17.0 Thai test clips whose sentences are not in its training data, and 7.28% of characters wrong on TVSpeech, natural speech from video dense with jargon.
- Speakers: each segment is labelled
SPEAKER_00,SPEAKER_01and so on, with its start and end time, for up to 8 speakers a file. - Long recordings: up to 8 hours of audio and 250 MB a file. An 85-minute MP3 took 9.3 minutes through the API.
- Names and terms: send the names you expect in
vocabulary, up to 500. On 2,127 YouTube clips, names were written exactly as given 84% of the time with the list, against 72% without it. - Thai, English and Chinese on the same endpoint, with
language_code.
Pro v2 is free until 30 November 2026. From 1 December 2026 it costs 0.003 IC a second of audio, billed by the second: 10.8 IC an hour, about 13.5 THB. From the same day, Transcribe plans give 18, 70 or 320 hours of audio a month for 199, 690 or 2,990 THB before VAT.
Base and Pro end on 15 November 2026. From that day, calls to their endpoints are answered by Pro v2 at Pro v2's price, with the same request and response shape, so no code change is needed. The Thai Speech-to-Text page lists what differs.
Chinda ASR Realtime: live speech
Chinda ASR Realtime takes Thai audio over a WebSocket, wss://api.iapp.co.th/v3/store/speech/speech-to-text/realtime-v2, and sends the text back while it arrives.
- The sentence so far: about every 0.7 seconds while someone speaks.
- The finished sentence: a median 0.66 seconds after the speaker pauses, 90% of them within 1 second.
- Accuracy: 12.95% of characters wrong on TVSpeech when streamed, against 12.41% for the same model on each whole clip. Pro v2 gets 7.28% on the same clips, so send recordings to Pro v2.
- Audio: Thai speech as 16 kHz 16-bit mono PCM, at the speed it is spoken, up to 4 hours a session.
Realtime is free during the launch. The earlier real-time endpoints, /v3/store/speech/speech-to-text/realtime, /asrx and /v1/audio/stt/rt, now run on it with the protocol they had; new integrations should use realtime-v2.
Try them
Both pages have a try-it for signed-in users: upload a recording to Pro v2, or speak into your microphone for Realtime. A new account comes with 50 free credits. For questions or volume, contact sale@iapp.co.th or 02-124-4041.