/v1/audio/transcriptions

Overview

Turns audio into text. Returns the full text, timestamped segments and words and, when you ask for it, who is speaking at each moment (diarization) in interviews, calls and meetings. Response formats json, verbose_json, diarized_json, text, srt and vtt.

Available models

Four literals on the same endpoint. Canary and Parakeet are the recommended ones: they transcribe real-world audio better, accept hours-long files without client-side chunking and separate speakers. Whisper stays available for compatibility with integrations that already use it.

Performance in Spanish and English versus Whisper

Measured by Tessera in October 2026 on the same GPU with the same normalisation (lower case, no punctuation, no filler words). FLEURS is short read sentences: the figure is WER (wrong words; lower is better). Real audio is long conversations: Spanish calls of 8 and 41 minutes and English interviews of 5 to 79 minutes with 2 or 3 speakers. There the figure is the difference against a reference transcript.

Endpoint and parameters

POST https://api.tesseraai.cloud/v1/audio/transcriptions with multipart/form-data.

Request

curl https://api.tesseraai.cloud/v1/audio/transcriptions \
  -H "Authorization: Bearer $TESSERA_API_KEY" \
  -F "[email protected]" \
  -F "model=nvidia/parakeet-tdt-0.6b-v3" \
  -F "language=en" \
  -F "response_format=verbose_json"

# Python (official OpenAI SDK)
import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.tesseraai.cloud/v1",
    api_key=os.environ["TESSERA_API_KEY"],
)
with open("call.mp3", "rb") as f:
    r = client.audio.transcriptions.create(
        model="nvidia/parakeet-tdt-0.6b-v3",
        file=f,
        language="en",
    )
print(r.text)

Response

With json it returns {"text": "..."}. With verbose_json it adds duration, language (the one you sent in the request), segments[] and words[] with start and end in seconds, and usage with the seconds of audio.

{
  "task": "transcribe",
  "language": "es",
  "duration": 85.07,
  "text": "Hola, gracias por venir. ¿Has tenido problemas para encontrar la oficina? ...",
  "segments": [
    {
      "id": 0,
      "start": 0.48,
      "end": 4.56,
      "text": "Hola, gracias por venir. ¿Has tenido problemas para encontrar la oficina?"
    }
  ],
  "words": [
    {"word": "Hola,", "start": 0.48, "end": 0.8},
    {"word": "gracias", "start": 0.88, "end": 1.2}
  ],
  "usage": {"type": "duration", "seconds": 85.07}
}

Multiple speakers (diarization)

Canary and Parakeet can tell who is speaking at each moment: interviews, calls, meetings or podcasts. It is off by default, but there is nothing to configure on your account and no extra cost: you turn it on per request. Two ways: add diarize=true (works with json, verbose_json, srt and vtt), or ask for response_format=diarized_json, which turns it on by itself and answers in OpenAI's format.

Turning diarization on

# cURL: transcript with speakers
curl https://api.tesseraai.cloud/v1/audio/transcriptions \
  -H "Authorization: Bearer $TESSERA_API_KEY" \
  -F "[email protected]" \
  -F "model=nvidia/canary-1b-v2" \
  -F "language=en" \
  -F "response_format=verbose_json" \
  -F "diarize=true" \
  -F "num_speakers=2"

# Python: Tessera-specific fields go in extra_body
with open("meeting.mp3", "rb") as f:
    r = client.audio.transcriptions.create(
        model="nvidia/canary-1b-v2",
        file=f,
        language="en",
        response_format="verbose_json",
        extra_body={"diarize": True, "num_speakers": 2},
    )
for s in r.segments:
    print(s.speaker, s.start, s.text)

# Python: OpenAI's diarized_json format (turns diarization on by itself)
with open("meeting.mp3", "rb") as f:
    r = client.audio.transcriptions.create(
        model="nvidia/parakeet-tdt-0.6b-v3",
        file=f,
        response_format="diarized_json",
    )
for s in r.segments:
    print(s.speaker, s.start, s.text)

Where the speakers appear

Speakers are labelled A, B, C… in order of appearance. Labels are consistent within one file but don't identify the person: the A in one call is not the A in another.

{
  "task": "transcribe",
  "duration": 85.07,
  "text": "Hola, gracias por venir. Has tenido problemas para encontrar la oficina. Ninguno, ...",
  "segments": [
    {
      "type": "transcript.text.segment",
      "id": 0,
      "speaker": "A",
      "start": 0.32,
      "end": 4.96,
      "text": "Hola, gracias por venir. Has tenido problemas para encontrar la oficina."
    },
    {
      "type": "transcript.text.segment",
      "id": 1,
      "speaker": "B",
      "start": 5.36,
      "end": 10.16,
      "text": "Ninguno, el metro me deja justo enfrente, llegué con 10 minutos de margen."
    },
    {
      "type": "transcript.text.segment",
      "id": 2,
      "speaker": "A",
      "start": 10.4,
      "end": 14.8,
      "text": "Perfecto, cuéntame un poco sobre tu experiencia en atención al cliente."
    }
  ],
  "usage": {"type": "duration", "seconds": 85.07}
}

Whisper (compatibility)

whisper-large-v3 and whisper-large-v3-turbo remain available with the same contract as always, so existing integrations keep working. Use them for Catalan and other languages outside the 25 Canary and Parakeet cover, and for /v1/audio/translations (translation into English). For everything else we recommend Canary or Parakeet.

• Up to 25 MB per file. For audio longer than 10 minutes, split it into 5–10 minute chunks and join the text client-side.
• whisper-large-v3-turbo cuts the decoder from 32 to 4 layers: faster, with a small accuracy loss.
• No diarization.

Best practices

• Send language when you know it. With Canary it is required.
• Send the whole file: Canary and Parakeet split and stitch the audio server-side, up to 4 hours.
• For conversations, turn on diarize=true and send num_speakers if you know how many people speak.
• The text, srt and vtt formats arrive inside a JSON {"text": "..."}. With the OpenAI SDK you get that string: read it with json.loads(r)["text"].
• Clean audio gives better results than heavily compressed mp3 or lots of background noise.