whisper-s2t-reborn 1.4.0

A streamlined Speech-to-Text pipeline for Whisper using CTranslate2

WhisperS2T-Reborn ⚡

A Streamlined Speech-to-Text Pipeline for Whisper Models using CTranslate2

---

WhisperS2T-Reborn is a streamlined fork of the original WhisperS2T project, focused exclusively on the CTranslate2 backend for fast and efficient speech transcription.

What's Different from the Original?

This fork simplifies the original WhisperS2T by:

• Single Backend Focus: Removed TensorRT-LLM, HuggingFace, and OpenAI backends—CTranslate2 only
• Curated Model Selection: Uses optimized CTranslate2 whisper models from ctranslate2-4you on HuggingFace
• Cleaner Codebase: Streamlined architecture with reduced dependencies
• Simplified Setup: Easier installation without complex backend configurations

Features

• 🚀 Fast Inference: CTranslate2 backend provides excellent speed/accuracy tradeoff
• 🎙️ Built-in VAD: Integrated Voice Activity Detection using NeMo's Marblenet models
• 🎧 Flexible Audio Input: Handles both small and large audio files efficiently
• 🌐 Multi-language Support: Transcription and translation for 99+ languages
• ⏱️ Word-level Timestamps: Optional word alignment for precise timing
• 📝 Multiple Export Formats: Export to TXT, JSON, TSV, SRT, and VTT

Supported Models

Model	English-only	Multilingual
tiny	✅ tiny.en	✅ tiny
base	✅ base.en	✅ base
small	✅ small.en	✅ small
medium	✅ medium.en	✅ medium
large-v3	—	✅ large-v3
distil-small.en	✅	—
distil-medium.en	✅	—
distil-large-v3	—	✅

Installation

Prerequisites

Install FFmpeg for audio processing:

Ubuntu/Debian:

apt-get install -y libsndfile1 ffmpeg

macOS:

brew install ffmpeg

Conda (any platform):

conda install conda-forge::ffmpeg

Install WhisperS2T-Reborn

CPU only:

pip install whisper-s2t-reborn

With GPU support (recommended for faster inference):

pip install whisper-s2t-reborn[gpu]

Note: The [gpu] extra installs NVIDIA CUDA libraries required for GPU acceleration with CTranslate2. Requires an NVIDIA GPU with compatible drivers.

Quick Start

Basic Transcription

import whisper_s2t

# Load model (downloads automatically on first use)
model = whisper_s2t.load_model(model_identifier="large-v3")

# Transcribe with VAD
files = ['audio/sample.wav']
out = model.transcribe_with_vad(files,
                                lang_codes=['en'],
                                tasks=['transcribe'],
                                initial_prompts=[None],
                                batch_size=32)

print(out[0][0])
# {'text': 'Your transcribed text here...',
#  'avg_logprob': -0.25,
#  'no_speech_prob': 0.0001,
#  'start_time': 0.0,
#  'end_time': 24.8}

With Word Timestamps

model = whisper_s2t.load_model("large-v3", asr_options={'word_timestamps': True})

out = model.transcribe_with_vad(files,
                                lang_codes=['en'],
                                tasks=['transcribe'],
                                initial_prompts=[None],
                                batch_size=32)

Export Transcripts

from whisper_s2t import write_outputs

# Export to various formats
write_outputs(out, format='srt', save_dir='./output/')
write_outputs(out, format='vtt', save_dir='./output/')
write_outputs(out, format='json', save_dir='./output/')

Translation

# Translate non-English audio to English
out = model.transcribe_with_vad(files,
                                lang_codes=['fr'],  # Source language
                                tasks=['translate'],  # Translate to English
                                initial_prompts=[None],
                                batch_size=32)

Configuration Options

Model Loading Options

model = whisper_s2t.load_model(
    model_identifier="large-v3",  # Model name or path
    device="cuda",                 # "cuda" or "cpu"
    compute_type="float16",        # "float16", "float32", or "bfloat16"
    asr_options={
        'beam_size': 5,
        'word_timestamps': False,
        'repetition_penalty': 1.01,
    }
)

Transcription Options

out = model.transcribe_with_vad(
    files,
    lang_codes=['en'],           # Language codes for each file
    tasks=['transcribe'],        # 'transcribe' or 'translate'
    initial_prompts=[None],      # Optional prompts for each file
    batch_size=32                # Batch size for inference
)

Acknowledgements

• Original WhisperS2T: This project is a fork of WhisperS2T by Shashi Kant Gupta
• OpenAI Whisper: The foundational Whisper model
• CTranslate2: Fast inference engine for Transformer models
• NVIDIA NeMo: VAD models used in this pipeline