whisperx

What it is and what it does

WhisperX is a speech-to-text system that wraps OpenAI's Whisper model with improvements for timestamp accuracy and speaker identification. It uses batched inference via faster-whisper for speed, forced phoneme alignment via wav2vec2 for word-level timing precision, and pyannote-audio for speaker diarization—assigning speaker labels to segments of audio. The package includes voice activity detection to reduce hallucinations and supports multiple languages through language-specific alignment models.

It is designed for transcription workflows where accurate timing and speaker identity matter: meeting notes, podcast chapters, video subtitles, or research datasets. The main trade-off is a large dependency tree (torch, transformers, torchaudio, and others) and GPU memory requirements, though CPU-only and reduced-precision modes are available. Setup requires CUDA 12.8 for GPU use and a Hugging Face token for speaker diarization.

Use it for:

• Transcribe meeting recordings with speaker labels and accurate word timestamps for searchable meeting notes.
• Generate subtitle files (.srt) with precise timing and speaker identification for video content.
• Process podcast or interview audio to extract speaker segments and create chapter markers.
• Batch-process large audio datasets for research or content indexing with minimal GPU memory.
• Transcribe multilingual audio (supported languages: en, fr, de, es, it, and others via Hugging Face).

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

WhisperX performs fast automatic speech recognition with word-level timestamps and speaker diarization, using batched inference and forced phoneme alignment to improve accuracy over standard Whisper.

Yes, if you need word-level timestamps and speaker diarization for transcription. The package is actively maintained, permissively licensed, and popular (top 5000 PyPI). Install friction is low but dependency footprint is substantial (torch, transformers, etc.); ensure GPU memory or CPU fallback fits your deployment. No known vulnerabilities as of 2026-08-14.

Install

whisperx on PyPI

pip install whisperx

uv add whisperx

poetry add whisperx

Installing whisperx

Before you install

Low friction installation via pip; 14 runtime dependencies including torch, torchaudio, and transformers mean a substantial download footprint. Active maintenance with recent releases; last commit 2026-07-13. Requires CUDA 12.8 for GPU acceleration or CPU fallback available.

License in practice

BSD-2-Clause permissive license allows commercial use, modification, and distribution with minimal restrictions—suitable for most production and research deployments.

Quickstart

pip install whisperx

import whisperx

device = "cuda"
audio_file = "audio.mp3"
model = whisperx.load_model("large-v2", device, compute_type="float16")
audio = whisperx.load_audio(audio_file)
result = model.transcribe(audio, batch_size=16)
print(result["segments"])

CUDA 12.8 required for GPU acceleration; CPU mode available but slower. Requires ffmpeg and Rust for some dependencies. Hugging Face token needed to enable speaker diarization.

Verify before relying

• Whether the 70x realtime speed claim applies to all model sizes or only large-v2 with specific batch settings.
• Memory requirements for different model sizes and batch configurations beyond the <8GB mention for large-v2.
• Whether all languages listed as supported via Hugging Face have been tested or are community-contributed.

Package facts

Evidence: whisperx-3.8.6-py3-none-any.whl

Tags

More Artificial Intelligence packages

Further reading