Faster Whisper Distilled | Worker

This repository contains the Faster Whisper Worker for RunPod, optimized to run distilled Whisper models. The Worker is designed to process audio files using efficient distilled variants of Whisper that maintain high accuracy while offering improved performance. It supports transcription formatting, language translation, and more as part of the RunPod Workers collection.

Endpoint Docs

Model Inputs

Input	Type	Description
audio	Path	Audio file
audio_base64	str	Base64-encoded audio file
model	str	Choose a Whisper model. Choices: "distil-large-v2", "distil-large-v3". Default: "distil-large-v2"
transcription	str	Choose the format for the transcription. Choices: "plain_text", "formatted_text", "srt", "vtt". Default: "plain_text"
translate	bool	Translate the text to English when set to True. Default: False
translation	str	Choose the format for the translation. Choices: "plain_text", "formatted_text", "srt", "vtt". Default: "plain_text"
language	str	Language spoken in the audio, specify None to perform language detection. Default: None
temperature	float	Temperature to use for sampling. Default: 0
best_of	int	Number of candidates when sampling with non-zero temperature. Default: 5
beam_size	int	Number of beams in beam search, only applicable when temperature is zero. Default: 5
patience	float	Optional patience value to use in beam decoding. Default: None
length_penalty	float	Optional token length penalty coefficient (alpha). Default: None
suppress_tokens	str	Comma-separated list of token ids to suppress during sampling. Default: "-1"
initial_prompt	str	Optional text to provide as a prompt for the first window. Default: None
condition_on_previous_text	bool	If True, provide the previous output of the model as a prompt for the next window. Default: True
temperature_increment_on_fallback	float	Temperature to increase when falling back when the decoding fails. Default: 0.2
compression_ratio_threshold	float	If the gzip compression ratio is higher than this value, treat the decoding as failed. Default: 2.4
logprob_threshold	float	If the average log probability is lower than this value, treat the decoding as failed. Default: -1.0
no_speech_threshold	float	If the probability of the token is higher than this value, consider the segment as silence. Default: 0.6
enable_vad	bool	If True, use the voice activity detection (VAD) to filter out parts of the audio without speech. This step is using the Silero VAD model. Default: False
word_timestamps	bool	If True, include word timestamps in the output. Default: False

Test Inputs

The following inputs can be used for testing the model:

{
    "input": {
        "audio": "https://github.com/runpod-workers/sample-inputs/raw/main/audio/gettysburg.wav"
    }
}

Sample output

{
    "segments": [
        {
            "id": 1,
            "seek": 106,
            "start": 0.11,
            "end": 3.11,
            "text": " Hola mundo.",
            "tokens": [
                50364,
                22637,
                7968,
                13,
                50514
            ],
            "temperature": 0.1,
            "avg_logprob": -0.8348079785480325,
            "compression_ratio": 0.5789473684210527,
            "no_speech_prob": 0.1453857421875
        }
    ],
    "detected_language": "es",
    "transcription": "Hola mundo.",
    "translation": null,
    "device": "cuda",
    "model": "distil-large-v2",
    "translation_time": 0.7796223163604736
}