Word timestamps. Speaker labels. No setup.
ACTUAL WHISPERX OUTPUT
Two short synthetic recordings, processed by WhisperX. Listen, inspect timestamps, and download the original results. These demos show how the tool works; they are not real-world accuracy tests.
0.05s · Single voice
Welcome to this WhisperX example.
2.73s · Single voice
Upload your audio, review the words as you listen, and download a transcript or subtitles.
Synthetic source audio and unedited model output. To try word navigation, editing, and speaker names, explore the editor.
0.09s · Voice 01
How do I run WhisperX without installing it?
3.20s · Voice 00
Upload an audio file, choose your language, and start transcription.
Synthetic source audio and unedited model output. To try word navigation, editing, and speaker names, explore the editor.
HUMAN SPEECH · OPEN TEST RECORDINGS
Two continuous excerpts from real meetings, processed by WhisperX. Hear the source, inspect the unedited result, and compare it with the AMI corpus reference. These small examples show both useful output and mistakes to check; they are not an accuracy benchmark.
Natural English speech · AMI IN1014 · Unedited transcript
Original WhisperX output · Click a passage to listen
Compared with the corpus reference transcript; not a full listening or timing audit.
Source: AMI Meeting Corpus, The AMI Consortium / University of Edinburgh, CC BY 4.0. Meeting IN1014, 1050.22–1112.23 seconds. Each is one continuous excerpt from a larger meeting, with no denoising or stitched speech. No endorsement implied.
Processed 2026-10-15 (UTC) · WhisperX large-v3 via Replicate.
English transcription, word alignment and automatic speaker detection enabled; batch size 32. Speaker count was not forced.
Run through the same production provider adapter and normalizer, outside the account/upload workflow. Model text and speaker labels were not manually corrected.
Pinned provider runtime (the exact upstream package build is not independently verified):
655845d6190ef70573c669245f245892cd039df4b880a1e3a65852c09252f5ccNatural English speech · AMI EN2009d · Unedited transcript
Original WhisperX output · Click a passage to listen
Compared with the corpus reference transcript; not a full listening or timing audit.
Source: AMI Meeting Corpus, The AMI Consortium / University of Edinburgh, CC BY 4.0. Meeting EN2009d, 4926.96–5023.53 seconds. Each is one continuous excerpt from a larger meeting, with no denoising or stitched speech. No endorsement implied.
Processed 2026-10-15 (UTC) · WhisperX large-v3 via Replicate.
English transcription, word alignment and automatic speaker detection enabled; batch size 32. Speaker count was not forced.
Run through the same production provider adapter and normalizer, outside the account/upload workflow. Model text and speaker labels were not manually corrected.
Pinned provider runtime (the exact upstream package build is not independently verified):
655845d6190ef70573c669245f245892cd039df4b880a1e3a65852c09252f5ccRecordings and reference annotations: The AMI Consortium / University of Edinburgh, AMI Meeting Corpus, CC BY 4.0. Both clips are excerpts from larger multi-person meetings. No endorsement implied.
WHAT IS WHISPERX?
WhisperX is an open source speech recognition pipeline. It combines transcription with word alignment and optional speaker diarization, so the result can tell you what was said, when it was said, and which voice said it.
WhisperX.app is an independent web app that runs that pipeline for you, without local installation. Bring your own audio or video, confirm the duration quote, and review the result in a browser editor. You can listen beside the transcript, correct text, name speakers, and keep the original machine output alongside your edits. English is available on the free trial. Unlock supported languages with paid access.
If you already know you want WhisperX, you can skip environment setup, model downloads, and GPU maintenance. Start with the public recordings below to check the output before uploading. This is an independently operated service, with no affiliation to the WhisperX authors or OpenAI. Explore the open source project.
WORD-LEVEL TIMESTAMPS
Click a word to hear it in the recording.
WhisperX adds an alignment stage after speech recognition to estimate word positions within the recording. In the editor, a timed word is a shortcut back to the audio. Use it to check a name, inspect a technical term, or listen to the sentence around a doubtful phrase.
Word-level JSON keeps the available times for your next workflow. Some words may have missing or estimated positions, and editing a sentence does not create fresh alignment. Keep the original JSON when you need to distinguish model output from reviewed text.
See how WhisperX differs from WhisperSPEAKER DIARIZATION
For an interview or conversation, enable speaker diarization to group voice turns. The two-voice demonstration shows actual labels returned by the model. Listen to a clear turn before renaming a group, then correct individual assignments where you hear a mismatch.
These groups belong to one recording. They do not identify people or match a voice across files. Overlap, short replies, and similar voices need closer review. If attribution is unclear, keep an unknown label instead of guessing.
Try the speaker diarization workflowHow do I run WhisperX without installing it?
Upload an audio file, choose your language, and start transcription.
Can I download Word timestamps and subtitles?
TRANSCRIPTS & SUBTITLES
ORIGINAL VTT CUE
WEBVTT
00:00:00.051 --> 00:00:02.315
Welcome to this WhisperX example.Download TXT for reading and writing, JSON for structured timing and speaker data, SRT for subtitle import, or VTT for a web player. All four formats are included. Choose either the original output or your edited version; downloading an existing result does not use more processing minutes.
A subtitle file accompanies your original media. Import it into your destination player or video editor and check the first, middle, and last cues. This release exports captions; it does not burn them into a new video, translate speech, or summarize the recording.
Create and review SRT or VTT subtitlesTHREE SIMPLE STEPS
The defaults handle the processing setup. You decide what to upload, whether to label speakers, and which output to keep.
Upload your own recording. Choose the language spoken in your audio or video, with speaker labels when you need them.
Confirm the duration quote and run WhisperX. Your task continues in your account after processing starts.
Listen, correct, and download. Keep the original result or export your reviewed transcript.
You need an account to process your own file. Minutes are reserved before processing and charged when the result is successfully delivered; a failed task releases its hold. Public sample playback, edits, and downloads never start a GPU task. Follow the step-by-step guide.
HOSTED OR LOCAL?
The open source software is available to run yourself. The hosted service is useful when you want its output without maintaining the environment.
Upload, confirm, review, and export in one workspace.
Install the project and manage your own runtime and files.
We handle model access and external GPU processing.
Choose CPU or a compatible GPU and configure the models yourself.
Pay for processing minutes; review and exports are included.
Open source code is available under its license; you supply compute and storage.
English on the free trial; supported languages with paid access.
Configure the upstream project according to its own capabilities and documentation.
Prefer to run it yourself? Read the WhisperX installation guide. For code and model terms, see licensing.
BEFORE YOU UPLOAD
Upload MP3, WAV, M4A, FLAC, OGG, MP4, MOV, or WebM that you have permission to process. Results are private to your account; the service does not fetch YouTube or other external links.
The one-time trial includes 5 minutes, with files up to 100 MB and 5 minutes long. Paid accounts allow files up to 500 MB and 60 minutes. Clear speech helps; every automatic transcript still needs review.
Trial results remain available for 7 days; paid results for 30 days. Download what you want to keep. Pack minutes last 12 months; subscription minutes last 30 days. Monthly and yearly plans renew until canceled.
CHOOSE YOUR WORKFLOW.
Listen, correct the transcript, rename speakers and export your result.
Review speaker turns, rename labels, and download.
Listen, correct the words, and export your text.
Review the transcript and export text or subtitles.
Download SRT or VTT, then check it in your destination player.
MINUTES, NOT COMMITMENTS
Try the workflow with 5 minutes, then choose a monthly or yearly plan, or a minute pack. Audio, video, speaker labels, and all four export formats use the same minute pricing.
$0USD
No card required. Once per verified account.
5 minutes to try
Trial minutes expire after 30 days
$9USD / month
Renews monthly. Cancel future renewal in Billing.
120 minutes every month
Monthly minutes expire after 30 days
$29USD / month
Renews monthly. Cancel future renewal in Billing.
600 minutes every month
Monthly minutes expire after 30 days
Buy processing minutes when you need them. No automatic renewal. Packs last 12 months and can be added alongside a subscription.
$9USD once
No subscription. Unused pack minutes expire after 12 months.
120 minutes to use as needed
$29USD once
No subscription. Unused pack minutes expire after 12 months.
600 minutes to use as needed
GET YOUR BEARINGS
The shortest route from a recording to a WhisperX result is to upload, run, and download. Use the editor when the recording needs a little more attention.
WHISPERX GUIDEIf you want control of the environment, run the open source project locally. This guide explains the setup decisions and links to the current upstream instructions.
WHISPERX GUIDEWhisperX builds a broader transcription pipeline around speech recognition. Its additional stages are most useful when your output needs timing and speaker information.
WHISPERX GUIDEA diarized transcript is easier to read when labels are checked and named. Start with clear turns, then examine the ambiguous parts of the conversation.
GOOD TO KNOW
English is available on the free trial. Unlock supported languages with paid access. Paid access supports English, Spanish, French, German, Italian, Portuguese, Chinese and Japanese. Transcription stays in the recording’s original language; it does not translate speech.
No. WhisperX.app is a web app: upload and review recordings in your browser while transcription runs on our external processing service. You do not need to install WhisperX locally. This is an independent service, not an official desktop or mobile app.
This website is an independent, unofficial hosted service. The open source project is maintained separately at m-bain/whisperX on GitHub.
No. The hosted service handles the processing environment. Upload your own file, confirm the charge, and run the task from your account.
An editable transcript, timing information, optional speaker labels, and downloads in TXT, JSON, SRT, or VTT. Review important words and speaker assignments before using the result.
Editing an existing result and downloading it again do not use more minutes. A new transcription run is a new task and requires a new confirmed charge.
Some words cannot be aligned reliably. Missing or uncertain timing is kept distinct from measured timing. Editing a sentence does not automatically align the new words.
Use a file upload. The service does not fetch YouTube videos, external recordings, or other media links.
LESS SETUP. MORE DONE.
Try the public sample, then bring your own recording. Your first account includes 5 minutes to try the workflow.
Run WhisperX