til-asr

Speech to text, with punctuation

Kazakh and Russian in one model. No need to name the language — it understands both and switches as the recording goes.

Word timings

See when every word was spoken

Add words=true and each word arrives with its start and end time. Times run through the whole recording: in an hour-long file the last word lands near second 3600.

{
  "duration_seconds": 2760,
  "text": "Сәлеметсіздер ме, білім алушылар…",
  "words": [
    {"word": "Сәлеметсіздер", "start": 0.42, "end": 1.08},
    {"word": "ме",            "start": 1.08, "end": 1.31}
  ]
}

Measured

How long it actually takes

Length	Size	Answer in
4 с	0,2 МБ	1–2 с
2,5 мин	9 МБ	12 с
10 мин	36 МБ	43 с
60 мин	220 МБ	4 мин 19 с

Measured on real recordings through the public API, not in a lab.

How to use it

One request

curl --max-time 900 https://router.tilqazyna.kz/v1/audio/transcriptions \
  -H "Authorization: Bearer ВАШ_КЛЮЧ" \
  -F "file=@zapis.m4a" \
  -F "model=til-asr" \
  -F "words=true"

Formats

Formats: wav, mp3, m4a, ogg, flac. We convert them ourselves.

Timeout

Raise your client timeout: the default thirty seconds only covers short recordings.

Long recordings

Long recordings are split on pauses between phrases. The answer comes back as one text.

Documentation