Transcription API

Overview

The Transcription API provides high-quality speech transcription. It supports automatic language detection, word-level timestamp alignment, speaker diarization, and speaker verification via reference audio samples.

Features include:

Automatic speech transcription with batch processing
Automatic language detection with configurable thresholds
Word-level timestamp alignment
Speaker diarization and speaker verification
Base64 audio input (no need to host files separately)
Configurable voice activity detection

Input Parameters

Parameter	Type	Required	Default	Description
`audio_file`	string	Yes	—	URL to an audio file, or base64-encoded audio data (optionally with `data:` URI prefix).
`language`	string	No	`null`	ISO language code (e.g. `en`, `fr`). If omitted, automatic detection is performed.
`language_detection_min_prob`	float	No	`0`	Minimum probability threshold for language detection.
`language_detection_max_tries`	int	No	`5`	Maximum number of language detection attempts.
`initial_prompt`	string	No	`null`	Optional prompt text for the first transcription window. Useful for vocabulary hints.
`batch_size`	int	No	`64`	Batch size for parallelized transcription. Lower values use less VRAM.
`temperature`	float	No	`0`	Sampling temperature. Higher values produce more random output.
`vad_onset`	float	No	`0.500`	Voice Activity Detection onset threshold.
`vad_offset`	float	No	`0.363`	Voice Activity Detection offset threshold.
`align_output`	bool	No	`false`	Enable word-level timestamp alignment.
`diarization`	bool	No	`false`	Enable speaker diarization (assigns speaker IDs to segments).
`min_speakers`	int	No	`null`	Minimum number of speakers. Only used when diarization is enabled.
`max_speakers`	int	No	`null`	Maximum number of speakers. Only used when diarization is enabled.
`speaker_verification`	bool	No	`false`	Enable speaker verification against reference samples.
`speaker_samples`	list	No	`[]`	List of `{"name", "url"}` objects for speaker verification.
`debug`	bool	No	`false`	Include compute/inference times and memory usage in the response.

Usage Examples

Basic transcription

{
  "input": {
    "audio_file": "https://example.com/audio/sample.wav"
  }
}

Base64 audio input

Send audio directly as base64-encoded data instead of a URL. Both raw base64 and data URI formats are supported:

{
  "input": {
    "audio_file": "data:audio/wav;base64,UklGRi..."
  }
}

Payload limits: 20 MB for /runsync, 10 MB for /run. Compress audio to MP3 or OGG before encoding for larger files.

Transcription with alignment

{
  "input": {
    "audio_file": "https://example.com/audio/sample.wav",
    "align_output": true,
    "batch_size": 32,
    "debug": true
  }
}

Full configuration with diarization

{
  "input": {
    "audio_file": "https://example.com/audio/sample.wav",
    "language": "en",
    "batch_size": 32,
    "temperature": 0.2,
    "align_output": true,
    "diarization": true,
    "min_speakers": 2,
    "max_speakers": 5,
    "debug": true
  }
}

Speaker verification

Provide reference audio for known speakers. There is no hard limit on the number of samples, but precision may decrease with many speakers.

{
  "input": {
    "audio_file": "https://example.com/audio/meeting.mp3",
    "language": "en",
    "batch_size": 32,
    "align_output": true,
    "diarization": true,
    "min_speakers": 2,
    "max_speakers": 5,
    "speaker_verification": true,
    "speaker_samples": [
      {"name": "Alice", "url": "https://example.com/alice.wav"},
      {"name": "Bob",   "url": "https://example.com/bob.wav"}
    ]
  }
}

Output Format

Without diarization

{
  "segments": [
    {
      "start": 0.0,
      "end": 2.5,
      "text": "Transcribed text segment one.",
      "words": [
        {"word": "Transcribed", "start": 0.1, "end": 0.7},
        {"word": "text",        "start": 0.8, "end": 1.2},
        {"word": "segment",     "start": 1.3, "end": 1.9},
        {"word": "one.",        "start": 2.0, "end": 2.4}
      ]
    }
  ],
  "detected_language": "en",
  "language_probability": 0.997
}

With diarization

{
  "segments": [
    {
      "start": 0.0,
      "end": 2.5,
      "text": "Hello, how are you?",
      "speaker": "SPEAKER_01",
      "words": [
        {"word": "Hello,", "start": 0.1, "end": 0.5, "speaker": "SPEAKER_01"},
        {"word": "how",    "start": 0.6, "end": 0.9, "speaker": "SPEAKER_01"},
        {"word": "are",    "start": 1.0, "end": 1.3, "speaker": "SPEAKER_01"},
        {"word": "you?",   "start": 1.4, "end": 1.8, "speaker": "SPEAKER_01"}
      ]
    },
    {
      "start": 2.5,
      "end": 5.0,
      "text": "I'm doing well, thanks.",
      "speaker": "SPEAKER_02",
      "words": [
        {"word": "I'm",    "start": 2.6, "end": 2.9, "speaker": "SPEAKER_02"},
        {"word": "doing",  "start": 3.0, "end": 3.4, "speaker": "SPEAKER_02"},
        {"word": "well,",  "start": 3.5, "end": 3.9, "speaker": "SPEAKER_02"},
        {"word": "thanks.","start": 4.0, "end": 4.5, "speaker": "SPEAKER_02"}
      ]
    }
  ],
  "detected_language": "en",
  "language_probability": 0.997,
  "speakers": {
    "SPEAKER_01": {"name": "Alice", "time": 2.5},
    "SPEAKER_02": {"name": "Bob",   "time": 2.5}
  }
}

Performance Notes

GPU memory — adjust batch_size based on available VRAM. Lower values reduce memory usage.
Processing time — enabling diarization and alignment increases processing time.
Language detection — may be less accurate on very short audio clips.