Faster Whisper

Faster Whisper template of Datastone Spirit Serverless.

Usage

Config

Minimal request body:

{
    "input": {
        "audio": "http://your-audio.wav"
    },
    "webhook": "http://your-backend-to-receive" 
}

or

{
    "input": {
        "audio_base64": "xxxrfsfsfsfs"
    },
    "webhook": "http://your-backend-to-receive" 
}

where "webhook" is for async request only.

Full request body:

{
    "input": {
        "audio": "http://your-audio.wav",
        "model": "base",
        "transcription": "plain_text",
        "translate": false,
        "language": null,
        "beam_size": 5,
        "best_of": 5,
        "patience": 1.0,
        "length_penalty": 1.0,
        "temperature": 0.0,
        "temperature_increment_on_fallback": 0.2,
        "initial_prompt": null,
        "condition_on_previous_text": true,
        "compression_ratio_threshold": 2.4,
        "log_prob_threshold": -1.0,
        "no_speech_threshold": 0.6,
        "enable_vad": false,
        "word_timestamps": false
    },
    "webhook": "http://your-backend-to-receive" 
}

If use webhook in async mode, the result will send to your webhook with query requestID=xxx-xxx&statusCode=200. You can find requestID from response of your async request.

Output format:

{
  "model": "base",
  "detected_language": "en",
  "device": "cpu",
  "segments": [
    {
      "id": 1,
      "seek": 1000,
      "start": 0,
      "end": 9.8,
      "text": " Four score and seven years ago, ...",
      "tokens": [
        50364,
        7451,
        ...
      ],
      "temperature": 0,
      "avg_logprob": -0.2194819552557809,
      "compression_ratio": 1.380952380952381,
      "no_speech_prob": 0.012501929886639118
    }
  ],
  "transcription": "Four score and seven years ago, ...",
  "translation": null,
  "word_timestamps": [
    {
      "word": " Four",
      "start": 0,
      "end": 0.6
    },
    {
      "word": " score",
      "start": 0.6,
      "end": 0.96
    },
    ...
  ]
}

argument	type	description
audio	str	URL of audio file
audio_base64	str	Base64 string of audio
model	str	Whisper model to use, available models: "tiny", "base", "small", "medium", "large-v1", "large-v2", "large-v3". Default: "base"
transcription	str	Type of output, available transcriptions: "plain_text", "formatted_text", "srt", "vtt". Default: "plain_text"
translate	bool	Translate to english or not, faster whisper only support translate to english now. Default: False
language	str	The language spoken in the audio. It should be a language code such as "en" or "fr". If not set, the language will be detected in the first 30 seconds of audio. Default: None
beam_size	int	Beam size to use for decoding. Default: 5
best_of	int	Number of candidates when sampling with non-zero temperature. Default: 5
patience	float	Beam search patience factor. Default: 1.0
length_penalty	float	Exponential length penalty constant. Default: 1.0
temperature	float	Temperature for sampling. Default 0.0
temperature_increment_on_fallback	float	Increment to temperature upon fallback. Increase temperature to 1.0 by temperature_increment_on_fallback. Default: 0.2. Means [0.0, 0.2, 0.4, 0.6, 0.8, 1.0]
initial_prompt	str	Optional text string or iterable of token ids to provide as a prompt for the first window. Default: None
condition_on_previous_text	bool	If True, the previous output of the model is provided as a prompt for the next window; disabling may make the text inconsistent across windows, but the model becomes less prone to getting stuck in a failure loop, such as repetition looping or timestamps going out of sync. Default: True
compression_ratio_threshold	float	If the gzip compression ratio is above this value, treat as failed. Default: 2.4
log_prob_threshold	float	If the average log probability over sampled tokens is below this value, treat as failed. Default: -1.0
no_speech_threshold	float	If the no_speech probability is higher than this value AND the average log probability over sampled tokens is below "log_prob_threshold", consider the segment as silent. Default: 0.6
enable_vad	bool	Enable the voice activity detection (VAD) to filter out parts of the audio without speech. This step is using the Silero VAD model https://github.com/snakers4/silero-vad. Default: False
word_timestamps	bool	If True, include word timestamps in the output. Default: False