Chinese Speech-to-Text

Turns a recording of Chinese speech into text, one timestamped segment per speaker turn. Two tiers take the same request and return the same response: Base at 1 IC per minute for recordings up to 10 hours, with 97.13% of characters right on the AISHELL-1 test set; Pro at 2 IC per minute for recordings up to 1 hour, which adds a correction step for speech with uncommon names and vocabulary.

POST/v3/store/speech/speech-to-text/base/zh

+1 endpoints

Try it

• Price1 ICper 60 seconds
• Measured97.13%
• Served89Kcalls

• v1.0
• Active
• Benchmark paper (PDF)

Input

OutputPOST /v3/store/speech/speech-to-text/base/zh

output

2 × text: Transcribed text from segment one. start: 0 end: 5.28 speaker: SPEAKER_00 segment: 0

audio_duration_in_seconds

60.5

uploaded_file_name

example.mp3

processing_time_in_seconds

12.34

use_asr_pro

false

asr_pro_is_used

false

iapp

seconds: 60.5

Every field, under ResponseOther endpoints: /zh see Request

Request

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

Response

200OK. The fields are explained under Response.

{
    "output": [
        {
            "text": "Transcribed text from segment one.",
            "start": 0.0,
            "end": 5.28,
            "speaker": "SPEAKER_00",
            "segment": 0
        },
        {
            "text": "Transcribed text from segment two.",
            "start": 5.28,
            "end": 10.56,
            "speaker": "SPEAKER_01",
            "segment": 1
        }
    ],
    "audio_duration_in_seconds": 60.5,
    "uploaded_file_name": "example.mp3",
    "processing_time_in_seconds": 12.34,
    "use_asr_pro": false,
    "asr_pro_is_used": false,
    "iapp": {
        "seconds": 60.5
    }
}

Request

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

Tier	Endpoint	Price	Longest recording
Base	POST /v3/store/speech/speech-to-text/base/zh	1 IC per 60 seconds	10 hours
Pro	POST /v3/store/speech/speech-to-text/pro/zh	2 IC per 60 seconds	1 hour

Headers

apikeystringrequired

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

Body

filefilerequired

The recording. MP3, WAV, M4A, AAC, AIF, CDA, FLAC, MID, OGG, WebM or WMA.

promptstring

The prompt template. Default base.

chunk_sizeinteger

Seconds of audio per processing chunk. Default 20.

The legacy paths /asr/v3/zh and /asr/v3/zh/pro answer at the same prices.

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

output.segmentinteger

The segment's index, from 0

audio_duration_in_secondsprocessing_time_in_secondsuploaded_file_name

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

iappobject

Billing

iapp.secondsnumber

The seconds billed

use_asr_proasr_pro_is_used

Whether the Pro model was asked for, and whether it was used

Status codes

Code	Meaning
400	No file, or an unsupported format
404	Task not found
500	Processing failed

Details

Measured

On the production service, 28 October 2026, against two public Mandarin test sets: AISHELL-1 (news sentences read by 20 speakers outside its training portion) and FLEURS (Wikipedia sentences read by non-professional speakers). Character error rate; lower is better.

Measure	Result
Base, AISHELL-1 test, 7,176 utterances	2.87%, so 97.13% of characters correct
Base, AISHELL-1 utterances with no character error	74.7%
Base, FLEURS Mandarin test, 945 utterances	10.58%
Base, the 654 FLEURS utterances written entirely in Chinese characters	3.43%
Pro against Base, the same 300 AISHELL-1 utterances	4.04% against 2.87%
Pro against Base, the same 150 FLEURS utterances	11.10% against 12.51%
Median time through the API, Base, clips of 4.6 s	1.25 s

Most of the FLEURS error comes from reference text the speakers do not say as written: foreign names repeated in Latin script, and numbers in digits. Pro lowered the error on FLEURS and raised it on the read news of AISHELL-1, where, of 300 answers, it wrote numbers as digits in 3, mixed in traditional characters in 2 and returned an English translation in 1. Choose Pro for recordings with uncommon names and vocabulary, and convert its text if your application accepts simplified characters only. Method and per-set results: the benchmark (PDF).

Limits

• Chinese speech. Up to 10 hours of audio per request on Base, 1 hour on Pro.

Data handling

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