Skip to main content
Hopper transcribes English audio with Nemotron ASR (nemotron-asr), as a file upload or as a live WebSocket stream.

Transcribe a file

POST /v1/audio/transcriptions takes a multipart/form-data upload and returns the transcript with word timestamps.
  • file (file, required) — PCM16 mono or WAV, up to 25 MB. Compressed audio is rejected; decode it first.
  • model (string, required) — nemotron-asr.
  • sample_rate (integer, optional, default 16000) — One of 8000, 16000, 22050, 24000, 44100, 48000.
The response carries the lowercase, unpunctuated transcript, one entry per word with its start and end offset in seconds, and the duration of the audio.
  • words[].text (string) — The word. It can carry a leading space; trim it.
  • words[].start, words[].end (number) — Offsets in seconds from the start of the audio.
  • words[].conf (number) — Reserved; currently always 0.

Transcribe a long recording

nemotron-3.5-transcribe transcribes recordings up to 1 hour in 40 languages, with punctuation and capitalization, on the same POST /v1/audio/transcriptions.
  • file (file, required) — WAV, MP3, M4A, FLAC or Ogg, up to 128 MB.
  • model (string, required) — nemotron-3.5-transcribe.
  • language (string, optional) — ISO 639-1 code: ar, bg, bn, cs, da, de, el, en, es, et, fi, fr, he, hi, hr, hu, id, it, ja, ko, lt, lv, mr, mt, nl, pl, pt, ro, ru, sk, sl, sv, ta, te, th, tl, tr, uk, vi, zh.
  • prompt (string, optional) — Instruction for the transcript, up to 4,000 characters.
  • response_format (string, optional, default json) — json returns the text; verbose_json adds timestamps in OpenAI’s shape.
  • timestamp_granularities[] (string, optional, repeatable) — With verbose_json: segment (default), word, or both.
  • speaker_labels (boolean, optional, default false) — Beta. Labels who is speaking: each turn in text starts with a tag such as <spk:0>, and with verbose_json every word and segment carries an integer speaker.
cURL
Billed per second of decoded audio; timestamps and speaker labels cost nothing extra.

Word timestamps and speakers

cURL
  • words[].word (string) — The word as written, with its punctuation.
  • words[].start, words[].end (number) — Seconds from the start of the file.
  • words[].speaker, segments[].speaker (integer) — Only with speaker_labels=true. Numbers tell speakers apart within one file; they are not identities.
  • segments[] — Consecutive words grouped at pauses, sentence ends and speaker changes, with id, start, end and text.
Timestamps add about 3 seconds per hour of audio. Speaker labels need the whole file in one pass, so a 1-hour recording takes about 30 seconds instead of about 10, and word times are less precise (typically within 0.5 s instead of 0.1 s). Labels can merge two similar voices, especially on phone audio.

Stream live audio

wss://api.withhopper.com/stt/websocket transcribes audio as it arrives. Authenticate with the Authorization header, or with the api_key query parameter from clients that cannot set headers.
  • api_key (string, optional) — Your key, when no Authorization header is sent.
  • model (string, optional, default nemotron-asr) — STT model id.
  • sample_rate (integer, optional, default 16000) — One of 8000, 16000, 22050, 24000, 44100, 48000.
  • turn_detection (string, optional, default balanced) — balanced, patient to wait longer before ending a turn, or responsive to end it sooner.
The server sends {"type": "Ready"} once, then you send binary PCM16 mono frames and receive transcript events. Send the text frame END to flush; the server replies with END when every final has been delivered.
  • text (string) — The turn’s transcript so far, lowercase and unpunctuated. It can carry a leading space.
  • is_final (boolean) — true on the last event of a turn; earlier events are partials that the next one replaces.
  • audio_start, audio_end (number) — Where the text sits in the stream, in seconds.
  • words (array) — Only on finals, in the same shape as the file response.
A session is billed on the audio bytes it receives, settled every 60 seconds and at close. If the balance reaches zero mid-session the server closes the socket with code 1008. Error codes