Faster Whisper — 4x Faster Speech-to-Text

What it is

Faster Whisper is a reimplementation of OpenAI's Whisper automatic speech recognition model using CTranslate2, a fast inference engine for Transformer models. It achieves up to 4x faster transcription compared to the original Whisper implementation while using less memory. The library supports GPU and CPU inference, 8-bit quantization, and all Whisper model sizes from tiny to large-v3.

Faster Whisper targets developers building transcription pipelines, voice-enabled applications, and audio processing workflows who need Whisper-quality results with better performance.

How it saves time or tokens

The original Whisper implementation uses PyTorch and processes audio at roughly real-time speed on CPU. Faster Whisper's CTranslate2 backend optimizes the computation graph, uses INT8 quantization, and batches operations more efficiently. The result is transcription that completes in a fraction of the audio duration, even on CPU.

For batch processing of audio files, the speed improvement means processing a day's worth of recordings in hours instead of a full day. GPU acceleration with float16 further reduces processing time.

How to use

1. Install Faster Whisper:

pip install faster-whisper

1. Transcribe an audio file:

from faster_whisper import WhisperModel

model = WhisperModel('large-v3', device='cuda', compute_type='float16')
segments, info = model.transcribe('audio.mp3')

for segment in segments:
    print(f'[{segment.start:.2f}s -> {segment.end:.2f}s] {segment.text}')

1. For CPU with INT8 quantization:

model = WhisperModel('large-v3', device='cpu', compute_type='int8')

Example

from faster_whisper import WhisperModel
import os

model = WhisperModel('medium', device='cpu', compute_type='int8')

# Transcribe with word-level timestamps
segments, info = model.transcribe(
    'meeting.wav',
    beam_size=5,
    word_timestamps=True,
    language='en'
)

print(f'Detected language: {info.language} (prob: {info.language_probability:.2f})')

for segment in segments:
    print(f'[{segment.start:.1f}s - {segment.end:.1f}s] {segment.text}')
    for word in segment.words:
        print(f'  {word.word} ({word.start:.2f}s - {word.end:.2f}s)')

This transcribes audio with word-level timestamps, useful for subtitle generation, keyword search indexing, and speaker diarization preprocessing.

Related on TokRepo

• AI voice tools -- Explore speech-to-text and voice synthesis tools
• Local LLM runners -- Run AI models privately on your hardware

Common pitfalls

• The large-v3 model requires approximately 3GB of VRAM with float16. If your GPU runs out of memory, use a smaller model (medium, small) or switch to INT8 quantization which halves memory usage.
• Audio files must be decoded before transcription. Faster Whisper handles common formats (MP3, WAV, FLAC) via ffmpeg. Make sure ffmpeg is installed on the system.
• Word-level timestamps add processing overhead. Disable word_timestamps when you only need segment-level timing to improve throughput.