Chinda ASR Realtime

Chinda ASR Realtime writes Thai speech down while people are still speaking, for live captions, call centres and voice input. Stream audio over a WebSocket: the sentence so far comes back about every 0.7 seconds, and each finished sentence a median 0.66 seconds after the speaker pauses, 90% of them within 1 second. Free during the launch.

A recording, or live speech

A recording? Use Chinda ASR Pro v2: it is more accurate, labels each speaker, takes a list of names and up to 8 hours a file. Need the text while people are talking? Use Realtime. On the TVSpeech test set, Realtime gets 12.95% of characters wrong; Pro v2 gets 7.28% wrong.

WSS/v3/store/speech/speech-to-text/realtime-v2Try it

• PriceFreeduring the launch
• Served130.9Kcalls

• Active

Sign in to run this demo.

50 credits free on sign-up. No card.

Create a free accountLog in

Connect

Open wss://api.iapp.co.th/v3/store/speech/speech-to-text/realtime-v2 with your API key in the apikey header. A browser cannot set headers on a WebSocket, so a web page may put the key in the URL as ?apikey=YOUR_API_KEY; connect from your server where you can, since a key in a web page can be copied.

1. Send one JSON text message with the settings below. Every field is optional.
2. The server answers {"type": "ready", "session_id": "..."}.
3. Send the audio as binary messages: 16-bit little-endian PCM, mono, 16 kHz, up to 1 second each (100 to 200 ms is a good size), at the speed it is recorded.
4. Read partial and final messages while you send.
5. Send the text message end. The server sends the last finals, then done, and closes the connection.

language_codestring

th, the default and the only language.

sample_rateinteger

16000, the default. Other rates are refused: resample on your side.

encodingstring

pcm_s16le, the default: 16-bit little-endian PCM, mono.

latency_msinteger

How much audio the service waits for before writing: 80, 320, 560 (the default) or 1120. Less is sooner and a little less accurate.

end_silence_msinteger

The pause that ends a sentence, 300 to 3000. Default 700.

{"language_code": "th", "sample_rate": 16000, "encoding": "pcm_s16le", "latency_ms": 560, "end_silence_ms": 700}

Messages

partialobject

The sentence so far. It changes as more audio arrives: show it, then replace it

partial.utteranceinteger

The sentence's number, from 0

partial.textstring

What has been said so far

partial.startnumber

When the sentence started, in seconds from the start of the session

finalobject

The finished sentence. It does not change again

final.utterancefinal.textfinal.start

As in partial

final.endnumber

When the sentence ended, in seconds from the start of the session

final.modelstring

The model version that answered, such as chinda-asr-realtime-v1

doneobject

After end, the last message before the close

done.audio_secondsnumber

The seconds of audio in the session

done.utterancesinteger

The number of sentences

A sentence ends when the speaker pauses for end_silence_ms, or after 30 seconds of speech without a pause. The messages below are the first sentence of the page's example recording, streamed on 9 November 2026.

{"type": "partial", "utterance": 0, "text": "สวัสดีค่ะชั้น", "start": 0.0}
{"type": "final", "utterance": 0, "text": "สวัสดีค่ะชั้นเสียงผู้ใหญ่ค่ะ", "start": 0.0, "end": 2.68, "model": "chinda-asr-realtime-v1"}
{"type": "done", "audio_seconds": 7.629, "utterances": 2}

Errors

An error arrives as {"type": "error", "code": "...", "message": "..."}. The connection then closes, except after BAD_AUDIO and BAD_MESSAGE. Opening the WebSocket without a valid key gets HTTP 401 before the upgrade.

Code	When
BAD_CONFIG	The first message is not valid JSON, or a setting has a value not listed above
BAD_AUDIO	A binary message with an odd number of bytes, or more than 1 second of audio; the message is skipped
BAD_MESSAGE	A text message other than end
TIMEOUT	No settings within 30 seconds of opening, or no audio for 30 seconds; finals for the audio already sent come first
TOO_LONG	The session reached 4 hours of audio
TOO_FAST	More than 2 minutes of audio is waiting: it is being sent much faster than real time. For a recording, use Pro v2
BUSY	Every live session is in use. Retry after a few seconds
ASR_FAILED	An internal error

Example

In Python, streaming a 16 kHz mono file at the speed it was recorded:

import asyncio, json
import numpy as np, soundfile as sf, websockets

async def main(path, key):
    audio, sr = sf.read(path, dtype="float32")          # a 16 kHz mono file, for the example
    pcm = (np.clip(audio, -1, 1) * 32767).astype("<i2").tobytes()
    url = "wss://api.iapp.co.th/v3/store/speech/speech-to-text/realtime-v2"
    async with websockets.connect(url, additional_headers={"apikey": key}) as ws:
        await ws.send(json.dumps({"language_code": "th"}))
        print(await ws.recv())                          # ready

        async def send():
            step = 3200                                 # 100 ms of 16-bit audio
            for i in range(0, len(pcm), step):
                await ws.send(pcm[i:i + step])
                await asyncio.sleep(0.1)                # real time
            await ws.send("end")

        async def receive():
            async for msg in ws:
                ev = json.loads(msg)
                if ev["type"] == "final":
                    print(f'{ev["start"]:7.2f}-{ev["end"]:7.2f}  {ev["text"]}')
                elif ev["type"] == "done":
                    break

        await asyncio.gather(send(), receive())

asyncio.run(main("meeting.wav", "YOUR_API_KEY"))

Details

Measured

On the production service, 16 live sessions streaming at real time, 9 November 2026. Character error rate, lower is better.

Measure	Result
The sentence so far	About every 0.7 s while someone speaks; the first about 1.1 s after speech starts
The finished sentence	A median 0.66 s after the speaker stops; 90% within 1 s
TVSpeech, 570 clips, streamed	12.95% (the same model on each whole clip: 12.41%; Pro v2: 7.28%)
10-minute recordings, streamed	7.99% (the same model on each whole recording: 8.72%)
Live sessions at once	Up to 20 across all customers at the launch; more get BUSY

Streamed text scores close to the same model reading each whole recording, so writing early costs little. Pro v2 is the more accurate model: send recordings to Pro v2.

Limits

• Thai speech, as 16 kHz 16-bit mono PCM.
• Up to 4 hours of audio a session.
• Audio at the speed it is recorded: more than 2 minutes waiting gets TOO_FAST.
• Not for live streams yet: a list of names to favour. It works for recordings on Pro v2.
• English words come back in Thai script.

Older endpoints

The earlier real-time endpoints, wss://api.iapp.co.th/v3/store/speech/speech-to-text/realtime, /asrx and /v1/audio/stt/rt, run on Chinda ASR Realtime since 9 November 2026 and keep the protocol they had: SegmentResult messages with is_final, then {"action": "stop"} and one SessionResult. They are free during the launch too. Two things changed:

• Thai only: transcribe_lang en or zh gets an error message, and the connection closes.
• The speaker fields are empty.

New integrations should use realtime-v2, as described on this page.

Data handling

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