WhisperX Online
EN
DeutschEnglishEspañolFrançaisPortuguês (Brasil)Русский日本語한국어繁體中文

WhisperX · online · free · private

WhisperX Online: transcribe audio with word-level timestamps

Drop an audio or video file and get a transcript with a timestamp for every word, speaker labels and SRT/VTT subtitles – no Python, no GPU setup, no Hugging Face token. Everything runs in your browser; your file is never uploaded.

Runs 100 % in your browser with WebGPU or WebAssembly. The model is downloaded once (40 MB to 800 MB) and then cached. Free: up to 15 minutes, word timestamps, all export formats.

  • No upload
  • Word timestamps
  • Speaker labels
  • SRT · VTT · JSON
  • 99 languages

What is WhisperX?

WhisperX is an open-source speech recognition pipeline built on OpenAI’s Whisper that adds accurate word-level timestamps and speaker diarization. It was developed by Max Bain and colleagues at the University of Oxford and published with the paper “WhisperX: Time-Accurate Speech Transcription of Long-Form Audio” (Interspeech 2023). According to its README, WhisperX combines three things: batched transcription with the faster-whisper backend, word alignment with a wav2vec2 phoneme model, and speaker labels from pyannote-audio. Voice activity detection cuts the audio first, which reduces Whisper’s hallucinations on silence.

The catch: WhisperX is a Python package. You need Python 3.10–3.13, ffmpeg, ideally an NVIDIA GPU with CUDA, and for speaker labels a Hugging Face token plus accepted model terms. The README only points to paid Replicate demos for people without a GPU. This page gives you the result without any of that.

How WhisperX Online works

  1. Drop a file. Audio (MP3, WAV, M4A, FLAC, OGG) or video (MP4, MOV, WEBM). Your browser extracts the audio track – nothing is uploaded.
  2. Choose model and language. Whisper tiny, base and small are free; large-v3-turbo is the most accurate. Language detection is automatic, or pick it for better results.
  3. Transcribe. The model is downloaded once and cached. With WebGPU (current Chrome, Edge, Safari) transcription is often faster than real time; without it, the CPU mode still works, just slower.
  4. Check and export. Click any word to jump to that moment in the player. Download TXT, SRT or VTT subtitles, or JSON in the WhisperX output format with segments, words, start, end and speaker.
How it differs from the original. WhisperX Online uses Whisper models compiled for the browser (transformers.js). Word timestamps come from Whisper’s cross-attention alignment instead of a separate wav2vec2 model, and speaker detection follows the pyannote 3 approach: the pyannote segmentation model finds speech, WeSpeaker voice embeddings recognise each voice, and similar voices are grouped into speakers. For long recordings with many speakers, the original WhisperX on a GPU is more precise – the “WhisperX command” tab gives you the exact command for your settings.

Free and Pro

FreePro (Webtools Pro)
ModelsWhisper tiny, base, small+ Whisper large-v3-turbo
File lengthup to 15 minuteslong recordings (limited by your device’s memory)
Word timestampsyesyes
Speaker detection–yes
Export TXT, SRT, VTT, JSONyesyes
PrivacyProcessing always in your browser – files are never uploaded

Pro is part of Webtools Pro by Answer the Web: one subscription for all web tools, including markitdown.de and crawl4ai.de.

WhisperX vs Whisper vs faster-whisper

Whisper (OpenAI)faster-whisperWhisperX
What it isThe original speech recognition model and Python packageRe-implementation of Whisper on CTranslate2Pipeline on top of faster-whisper
SpeedBaselineSeveral times faster, less memory (per its README)Batched inference, very fast on GPU
TimestampsSegment level, word level approximateWord timestamps availableWord timestamps via wav2vec2 forced alignment
SpeakersNoNoYes, with pyannote-audio
LicenseMITMITBSD-2-Clause

Rule of thumb: Whisper for simple transcripts, faster-whisper when speed on your own hardware matters, WhisperX when you need exact word timing (subtitles, karaoke, audio editing) or who-said-what.

Install WhisperX yourself

pip install whisperx
whisperx audio.mp3 --model large-v3 --language en --output_format srt

With speaker labels (needs a Hugging Face read token and accepting the terms of the pyannote diarization model):

whisperx audio.mp3 --model large-v3 --diarize --hf_token YOUR_HF_TOKEN

On a machine without NVIDIA GPU add --compute_type int8. The current release on PyPI is 3.8.6 (May 2026). Source, issues and documentation: github.com/m-bain/whisperX.

WhisperX: frequently asked questions

What does WhisperX do?
WhisperX transcribes speech with OpenAI’s Whisper and then aligns every word to the audio, so each word gets an exact start and end time. Optionally it labels speakers (diarization). The result is ideal for subtitles, searchable transcripts, interviews and podcasts.
What is the difference between Whisper and WhisperX?
Whisper produces the text with rough timestamps. WhisperX adds voice activity detection, faster batched inference, precise word-level alignment with a wav2vec2 model and speaker labels from pyannote. The words themselves come from Whisper in both cases.
How accurate is WhisperX?
Text accuracy depends on the Whisper model: large models are clearly better than tiny or base, especially for accents and noisy audio. WhisperX’s strength is timing – its paper reports more accurate word timestamps on long recordings than plain Whisper. Always proofread names and numbers.
Is WhisperX free?
Yes. WhisperX is open source under the BSD-2-Clause license and free to use, also commercially. You only need your own hardware. WhisperX Online is free for files up to 15 minutes with the tiny, base and small models.
Is my audio uploaded?
No. The file is decoded and transcribed inside your browser. Only the model files are downloaded from Hugging Face; your audio never leaves your device.
Is WhisperX better than faster-whisper?
WhisperX uses faster-whisper internally and adds word alignment and speaker labels. If you only need text, faster-whisper is enough; if you need exact word timing or speakers, WhisperX is the better choice.
Is this the official WhisperX website?
No. WhisperX Online is an independent interface by Answer the Web. The official project is github.com/m-bain/whisperX.
Which browsers are supported?
Current Chrome, Edge and Safari use WebGPU for fast transcription. Firefox and older browsers run in CPU mode, which works but is slower. On phones, use the tiny or base model.