Chinda ASR Thai Speech-to-Text

Chinda ASR Pro v2 turns a recording of Thai speech into text, one timestamped segment per speaker turn, with each speaker labelled (SPEAKER_00, SPEAKER_01 and so on). It gets 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 takes up to 8 hours of audio per request. Names and terms you send in vocabulary come back spelled your way. The same endpoint also transcribes English and Chinese (language_code en or zh). For text while people are still talking, use OpenThai ASR Realtime.

Pro v2 price, and the end of Base and Pro

Pro v2 is free until 30 November 2026. From 1 December 2026, 0.003 IC per second of audio, billed by the second: 10.8 IC per hour of audio, about 13.5 THB. Or, from the same day, a monthly plan: Transcribe S, M and L give 18, 70 and 320 hours of audio a month for 199, 690 and 2,990 THB before VAT.

Base and Pro end on 15 November 2026. From that day, calls to …/speech-to-text/base, …/speech-to-text/pro and the legacy /asr/v3 and /asr/v3/pro are answered by Pro v2 at Pro v2's price, with the same request and the same response shape, so no code change is needed. What differs: Pro v2 takes language_code th, en and zh, and any other value gets 400; it takes up to 8 hours of audio and 250 MB per request; it writes numbers as words, English words in Thai script and colloquial words in standard spelling; it labels each speaker (SPEAKER_00, SPEAKER_01 and so on); it does not return avg_logprob; and it always sets use_asr_pro and asr_pro_is_used to true. The English and Chinese Base and Pro endpoints do not change.

POST/v3/store/speech/speech-to-text/pro-v2

+2 endpoints

Try it

Input

OutputPOST /v3/store/speech/speech-to-text/pro-v2

output: 2 × text: สวัสดีค่ะชั้นเสียงผู้ใหญ่ค่ะ start: 0 end: 3.4 speaker: SPEAKER_00 segment: 0
audio_duration_in_seconds: 7.63
uploaded_file_name: 2f3aeeb3-0cb6-42fb-b694-933b1a2c73f4.wav
processing_time_in_seconds: 1.041
use_asr_pro: true
asr_pro_is_used: true
iapp: seconds: 7.63

Every field, under ResponseOther endpoints: /base, /pro see Request

Request

curl -X POST https://api.iapp.co.th/v3/store/speech/speech-to-text/pro-v2 \
  -H "apikey: YOUR_API_KEY" \
  -F "file=@recording.wav"

Response

200OK. The fields are explained under Response.

{
  "output": [
    { "text": "สวัสดีค่ะชั้นเสียงผู้ใหญ่ค่ะ", "start": 0.0, "end": 3.4, "speaker": "SPEAKER_00", "segment": 0 },
    { "text": "สวัสดีครับผมเสียงผู้ใหญ่ผู้ชายครับ", "start": 3.4, "end": 7.63, "speaker": "SPEAKER_01", "segment": 1 }
  ],
  "audio_duration_in_seconds": 7.63,
  "uploaded_file_name": "2f3aeeb3-0cb6-42fb-b694-933b1a2c73f4.wav",
  "processing_time_in_seconds": 1.041,
  "use_asr_pro": true,
  "asr_pro_is_used": true,
  "iapp": { "seconds": 7.63 }
}

Monthly plans

Transcribe S18 hours a month199 THB a month, 11.06 THB per hour
or 1,990 THB a year, two months freeFrom 1 December 2026
Transcribe M70 hours a month690 THB a month, 9.86 THB per hour
or 6,900 THB a year, two months freeFrom 1 December 2026
Transcribe L320 hours a month2,990 THB a month, 9.34 THB per hour
or 29,900 THB a year, two months freeFrom 1 December 2026

These plans can be bought from 1 December 2026.

Prices before 7% VAT. Every Chinda ASR Pro v2 speech-to-text endpoint counts against the plan; beyond it, calls cost 10.8 credits per hour from your balance, the pay-as-you-go price. Renews by card each month or year; cancel any time, effective at the end of the period.

Request

multipart/form-data, one recording per request, to the endpoint of the tier you want.

Tier	Endpoint	Price
Base	`POST /v3/store/speech/speech-to-text/base`	1 IC per 60 seconds, until 14 November 2026
Pro	`POST /v3/store/speech/speech-to-text/pro`	2 IC per 60 seconds, until 14 November 2026
Pro v2	`POST /v3/store/speech/speech-to-text/pro-v2`	Free until 30 November 2026, then 0.003 IC per second (10.8 IC per hour)

Headers

apikeystringrequired

Your API key, created under API Keys in the control panel.

Body

filefilerequired

The recording. MP3, WAV, AAC, M4A or WebM; Pro v2 also reads OGG/Opus, FLAC and the sound of MP4 and other video files. Length and size per request are under Limits.

language_codestring

th, the default. Pro v2 also takes en and zh, and en-US, zh-CN, zh-TW, zh-HK and zh-Hant; the last three give Traditional characters. Any other value gets 400.

vocabularystring

Pro v2. Names and terms that may be said, spelled the way you want them written: people, companies, brands, products, places, jargon. Repeat the field, or send one with an item per line or between commas, or a JSON list; up to 500 items of up to 60 characters. See Names and terms.

speaker_labelsboolean

Pro v2. true, the default, labels each speaker; false skips that step, which is a little faster, and every segment is SPEAKER_00.

mai_yamokstring

Pro v2, Thai. How a repeated word is written: expand, the default, writes it out (จริงจริง), as transcripts of speech usually do; keep writes it with the repetition mark (จริงๆ), as standard written Thai does.

chunk_sizeinteger

Seconds of audio per processing chunk. Default 7. Pro v2 accepts it and ignores it.

The legacy paths /asr/v3 and /asr/v3/pro answer at the same prices until 14 November 2026. From 15 November 2026, Pro v2 answers every one of these paths.

Response

outputarray

The segments, in the order spoken

output.textstring

What was said

output.startnumber

When the segment starts, in seconds

output.endnumber

When it ends, in seconds

output.speakerstring

The speaker's label: SPEAKER_00, SPEAKER_01 and so on; every segment is SPEAKER_00 with speaker_labelsfalse

output.segmentinteger

The segment's index, from 0

output.avg_logprobnumber

How clearly the recogniser heard this segment: the mean log-probability of its tokens, 0 or below, closer to 0 is clearer. Measured: −0.04 on clean speech, −0.21 with light background noise, −0.30 with heavy noise. Below −0.15, have a person check the segment before you take dates, numbers or names from it.

audio_duration_in_secondsprocessing_time_in_secondsuploaded_file_name

The recording's length, the time taken, the file name

use_asr_proasr_pro_is_usedboolean

Whether the Pro step was asked for, and whether it was used: false on Base, true on Pro

iappobject

Billing

modelstring

Pro v2: the model version that answered, such as chinda-asr-pro-v2.2

Status codes

Code	Meaning
200	Also for a recording with no speech: `output` is an empty list
400	No file, an empty file, a file that cannot be read as audio, a `language_code` other than `th`, `en` or `zh`, a `vocabulary` over 500 items, with an item over 60 characters, or in JSON that is not a list of strings, or a `mai_yamok` other than `expand` or `keep`
401	No API key, or one that is not valid
404	Task not found
413	Audio longer, or an upload larger, than the limits under Limits
429	Over your API key's rate limit (Pro v2: 5 requests a second, 60 a minute, 3,000 an hour); see Limits
500	Processing failed (on a long recording after the 200 has started, see Long recordings)
503	Busy for a moment: retry after the number of seconds in the `Retry-After` header

Long recordings

Pro v2 takes about an eighth of the audio's length: an 85-minute MP3 took 9.3 minutes through the API. For a recording over 5 minutes it answers 200 and sends the headers at once, then one space every 20 seconds while it works, then the usual JSON. JSON parsers skip leading spaces, so client code does not change. Do not set a total timeout shorter than the processing time; a read timeout of 60 seconds or more is enough. If transcription fails after the 200 has started, the body is {"detail": {"code": "ASR_FAILED", "message": "...", "retryable": true, "request_id": "..."}}: check for detail and send the recording again.

Names and terms

Send the names and terms you expect in vocabulary, spelled the way you want them written. Where a reading the audio supports matches an item, Pro v2 prefers it, so a name that was not said is not written in. It helps most with names that have several spellings (ศรัณย์ or ศรัณ), company and product names, and jargon. The list goes with each request and adds about 7% to the processing time.

curl -X POST https://api.iapp.co.th/v3/store/speech/speech-to-text/pro-v2 \
  -H "apikey: YOUR_API_KEY" \
  -F "file=@meeting.mp3" \
  -F "vocabulary=ศรัณย์ วงศ์สวัสดิ์" \
  -F "vocabulary=ไอแอพพ์ เทคโนโลยี"

Details

Measured

On the production service, 28 October 2026, against three public Thai test sets: Common Voice 17.0 (everyday sentences read aloud), FLEURS (Wikipedia sentences read aloud) and TVSpeech (natural speech from video, dense with jargon). Character error rate, spaces ignored; lower is better.

Measure	Result
Base, Common Voice 17.0 Thai test, 11,042 clips	5.21%, so 94.79% of characters correct
Base, FLEURS Thai test, 1,021 clips	9.93%
Base, TVSpeech, 570 clips	18.50%
Pro against Base, the same 300 Common Voice clips	5.11% against 5.59%
Pro against Base, the same 80 TVSpeech clips	17.69% against 19.21%
Pro v2, Common Voice 17.0 Thai test, the 3,539 clips whose sentences are not in its training data	2.06%, so 97.94% of characters correct (Base on the same clips: 5.33%)
Pro v2, Common Voice 17.0 Thai test, all 11,042 clips	1.89%, so 98.11% of characters correct
Pro v2 against Base, FLEURS Thai test, 1,021 clips	9.43% against 9.93%
Pro v2 against Base, TVSpeech, 570 clips	7.28% against 18.50%
Pro v2, English: LibriSpeech test-clean, 300 clips / FLEURS, 200 clips (8 Nov 2026)	Word error rate 1.61% / 3.87%
Pro v2, Chinese (Mandarin): FLEURS, 200 clips (8 Nov 2026)	Character error rate 7.07%
Pro v2 with `vocabulary`, GigaSpeech 2 dev, 2,127 YouTube clips (9 Nov 2026)	Names written exactly as given: 72% without the list, 84% with it; error rate 5.05% without, 4.90% with
Pro v2 with `vocabulary`, TVSpeech, 570 clips (9 Nov 2026)	Names written exactly as given: 67% without the list, 74% with it; error rate unchanged
Pro v2 with a list of names not in the recordings, TVSpeech (9 Nov 2026)	Names written exactly: 67% to 68%; error rate 0.08 points higher; none of the listed names written in
Pro v2 through the API, an 85-minute MP3	9.3 minutes
Median time through the API, Base, clips of 4.4 s / 26.1 s	0.68 s / 2.13 s
Median time through the API, Pro, clips of 4.3 s / 27.5 s	4.52 s / 6.01 s

1,186 of the Common Voice test clips (10.7%) repeat a sentence from the Common Voice 11.0 training or validation splits, which are in the service's training data; on the other 9,856 clips Base measures 5.27%. Pro lowers the error most on natural speech with names and jargon; on clear read speech its difference from Base is within chance. Method and per-set results: the benchmark (PDF).

Pro v2 was measured on the same sets with the same scorer, release v2.2.6 of 9 November 2026, which writes colloquial words in standard spelling. On the 3,539 Common Voice 17.0 Thai test clips whose sentence and recording are in none of its training data, Pro v2 transcribes 97.94% of characters correctly (character error rate 2.06%; 97.95% before the spelling change). On all 11,042 clips: 98.11%. The other 7,503 clips are in its training data; the two groups differ by 0.25 points. Pro v2 writes numbers as words, which counts as an error where a reference writes digits: with numbers spelled out on both sides, FLEURS measures 7.01% for Pro v2 against 8.63% for Base. On TVSpeech, Pro v2 is better than Base by 11.22 points (95% CI 8.09 to 13.53, resampling the 18 source videos).

For the vocabulary rows, each recording's list held the names in its own reference transcript, about 23 a recording; the last row gave each recording the names of other recordings instead.

Noisy recordings

Pro corrects the recognised text in context. Its own errors read fluently, such as an added or dropped connective or a dropped word, and on a noisy recording it can turn words that were not heard clearly into fluent text that says something else, such as a different month. avg_logprob describes the audio, not the corrected text, and is the same on both tiers, so use it to find the segments to review: in a test sentence mixed with light noise, the segment with the wrong month scored −0.21. English technical terms spoken with Thai pronunciation, such as Login or OAuth, can come back as garbled transliterations even at a confident score (−0.09).

Limits

Base and Pro: Thai speech, up to 30 minutes of audio per request.
Pro v2: Thai, English and Chinese, up to 8 hours of audio per request and 250 MB per upload; above either, 413. A 16 kHz 16-bit WAV reaches 250 MB at about 2 hours 10 minutes, so send long recordings as MP3, M4A or Opus (8 hours of 64 kbps MP3 is about 230 MB). Up to 8 speakers per file.
Pro v2 writes Thai numbers as words, English words in Thai script and colloquial words in standard spelling (อะ, ไหม, เขา, not อ่ะ, มั้ย, เค้า), without punctuation; a repeated word is written out (จริงจริง) unless mai_yamok is keep, and a word you send in vocabulary keeps your spelling; English and Chinese come with punctuation and numbers as digits, Chinese in Simplified characters, or Traditional for zh-TW, zh-HK and zh-Hant.
Pro v2 rate limits: per API key, 5 requests per second, 60 per minute and 3,000 per hour. Over the limit you get 429; the X-RateLimit-Remaining-Second, -Minute and -Hour response headers show what is left.

Data handling

Consent and notice under PDPA and GDPR remain with you as the data controller.

Monthly plans​

Request​

Response​

Details​

Monthly plans

Request

Response

Details