Audio Subtitler

Convert audio files to subtitles (VTT, SRT) using Faster-Whisper.

Features

• 🚀 Full Faster-Whisper support - All features and parameters from faster-whisper
• 📝 Multiple formats - VTT (WebVTT) and SRT subtitle output
• 🎯 Smart auto-detection - Automatically detects format from file extension
• 🌍 Multi-language - Supports 100+ languages with auto-detection
• ⚡ GPU acceleration - CUDA support for faster transcription
• 🎙️ Voice Activity Detection - Automatically removes silence
• 💻 Simple APIs - Easy-to-use CLI and Python API
• 🐳 Docker GPU support - Ready for serverless deployment

Installation

pip install audio-subtitler

Optional dependencies:

pip install audio-subtitler[runpod]  # For RunPod serverless
pip install audio-subtitler[dev]     # For development

Quick Start

CLI

# Auto-detect format from file extension (recommended)
audiosubtitler input.mp3 -o output.vtt
audiosubtitler input.mp3 -o output.srt

# Specify options
audiosubtitler input.mp3 -o output.vtt --model large-v3 --language en --device cuda

# Output to stdout
audiosubtitler input.mp3 --format srt > output.srt

# Use shorter command
audiosub input.mp3 -o output.vtt

# Hint punctuation style (helps large-v3 and others output more periods, commas, etc.)
audiosubtitler input.mp3 -o output.vtt --initial-prompt "Hello. How are you? Thanks."

Python API

from src import AudioSubtitler

# Initialize
converter = AudioSubtitler(
    model_size_or_path="base",
    device="cpu",
    compute_type="int8"
)

# Transcribe (returns subtitle string directly)
vtt = converter.transcribe("audio.mp3", format="vtt", language="en")
print(vtt)  # "WEBVTT\n\n00:00:00.000 --> ..."

srt = converter.transcribe("audio.mp3", format="srt")

# Better punctuation: pass a short punctuated phrase as a hint
vtt = converter.transcribe("audio.mp3", format="vtt", initial_prompt="Hello. How are you? Thanks.")

Getting better punctuation (for subtitles)

Larger Whisper models (e.g. large-v3) are more accurate on words but often output longer segments with less punctuation. Parameters that affect punctuation and segment boundaries:

Example for more punctuation:

audiosubtitler input.mp3 -o output.vtt --vad-silence-duration-ms 400 --patience 0.5

2. Post-process with a punctuation restoration model
   For English, you can run the transcript through a dedicated punctuation model (e.g. rpunct or speechbox) and then regenerate VTT/SRT from the punctuated text if your tool supports it.

API Reference

AudioSubtitler

Constructor: AudioSubtitler(**kwargs)

Accepts all faster-whisper WhisperModel parameters:

• model_size_or_path: Model name (tiny, base, small, medium, large, large-v3) or path
• device: "cpu", "cuda", or "auto"
• compute_type: "int8", "int8_float16", "int16", "float16", "float32"
• cpu_threads, num_workers, download_root, local_files_only, etc.

Method: transcribe(audio, format="vtt", **kwargs)

Parameters:

• audio: File path (str), file object (BinaryIO), or numpy array
• format: "vtt" or "srt" (default: "vtt")
• **kwargs: All faster-whisper transcribe parameters
  • language, beam_size, vad_parameters, word_timestamps, etc.

Returns: str — The subtitle content (VTT, SRT, or JSON string depending on format).

Docker (GPU only)

docker-compose -f docker-compose-gpu.yml up

RunPod serverless

Input (in the job input):

{
  "audio": "<base64_encoded_audio>",
  "format": "vtt"
}

• audio: required, base64-encoded audio bytes
• format: optional, "vtt" (default), "srt", or "json"

Output: the handler returns the subtitle string directly (no wrapper). RunPod puts it in the job result’s output field, so the response body looks like:

{
  "delayTime": 1119,
  "executionTime": 499,
  "id": "...",
  "output": "WEBVTT\n\n00:00:00.000 --> 00:00:00.280\nHello\n\n...",
  "status": "COMPLETED",
  "workerId": "..."
}

Use response["output"] to get the VTT/SRT/JSON string.

Errors: the handler raises exceptions (e.g. no audio, invalid base64, transcription failure). RunPod surfaces these in its error response.

Output Examples

VTT:

WEBVTT

00:00:00.000 --> 00:00:03.500
Hello, this is a test transcription.

00:00:03.500 --> 00:00:07.200
The audio is converted to text with timestamps.

SRT:

1
00:00:00,000 --> 00:00:03,500
Hello, this is a test transcription.

2
00:00:03,500 --> 00:00:07,200
The audio is converted to text with timestamps.

Environment Variables

License

MIT License - see LICENSE file for details.