WhisperX vs WhisperAI

You want a transcript that says who spoke, and someone pointed you at WhisperX, the open-source Python pipeline built on Whisper. We followed its README on a Mac. The first install failed, the Mac's GPU wasn't an option, and --diarize transcribed the whole clip, then stopped at a locked model and wrote no files. You don't need any of that to get speaker labels. In WhisperAI, they're a switch you turn on when you upload.

By Eman Khoubian, Founder & CEO

WhisperX vs WhisperAI for transcription

From the WhisperX README and PyPI, our own test, and WhisperAI's plans. As of November 3, 2026.
	WhisperAI	WhisperX
What you install	Nothing. It's a website, and there's an iPhone app.	Python below 3.14, ffmpeg, and 102 packages on our Mac
What it runs on	Any computer with a browser	Your machine: an NVIDIA GPU with CUDA, or the CPU.
Speaker labels	A switch at upload, on every plan. Labels read Speaker 1 until you rename them.	The --diarize flag, a Hugging Face token and a gated model. Labels read SPEAKER_00.
Timestamps	Each sentence in the web app, each word through the Developer API	Each word. Default models cover five languages, and words such as "2014." or "£13.60" get no timing.
Where the audio goes	Uploaded to our cloud, with TLS 1.3 in transit and AES-256 in storage	It stays on your machine
Files you get	PDF, Word, TXT, SRT, VTT and JSON	JSON, SRT, VTT, TSV and TXT
Price	Free for 5 minutes a month. Premium is $14.99 for 120 minutes. Business Pro is $24.99 with no minute limit.	Free code. You supply the machine.
When it breaks	Write to our team	GitHub issues. The README lists no support promise.

WhisperX download and install: what happened on our Mac

We started on an Apple M5 Pro with 64 GB of memory, and Python, ffmpeg and uv were already on it. The Python was Homebrew's current one, 3.15. Both install lines in the README fail on it. Here's pip:

$ pip install whisperx
ERROR: No matching distribution found for ctranslate2==4.4.0

That error names a dependency, not Python, and the README's install section doesn't give a Python version. PyPI does: below 3.14. On 3.13, this worked:

$ uv venv --python 3.13 venv
$ source venv/bin/activate
$ uv pip install whisperx
$ whisperx clip.wav --model small --device cpu --compute_type int8

The install and its models came to about 1.9 GB on disk. A 5-minute clip came back with a time for each word in the JSON:

{"word": "LibriVox", "start": 4.85, "end": 5.39, "score": 0.816}

Memory peaked at about 4.2 GB in that one run, on the small model. The Mac's GPU is out: the README's line for Mac is --device cpu, and --device mps stops with ValueError: unsupported device mps.

A developer has done worse before breakfast. If you came here to get a transcript, that's a lot of setup, and we still had no speaker labels.

WhisperX transcriber with speaker labels: what it takes

Speaker labels are called diarization in WhisperX, and they're the step that sends you off to another website.

WhisperAI

Sign up
Upload the file
Turn on Speaker Labels
Rename Speaker 1 and Speaker 2

WhisperX

Have Python below 3.14 and ffmpegWe did
Install whisperxWe did, on Python 3.13
Make a Hugging Face accountWe stopped here
Accept the conditions on pyannote's model pageNot done
Create a read access tokenNot done
Run whisperx with --diarize and --hf_tokenNot run

The steps to a transcript with speaker labels. WhisperX's come from its README and PyPI, and the notes show how far we got on November 3, 2026.

We ran --diarize with no token, the way the README's example is written. It printed a warning that the token was missing, carried on, transcribed and aligned the whole clip, and then stopped:

Cannot access gated repo for url https://huggingface.co/pyannote/
speaker-diarization-community-1/resolve/main/config.yaml.

The output folder was empty, so the transcript it had just made was gone. The fix is steps 3 to 5 in the list above: pyannote's page asks you to agree to share your Hugging Face details, and says it "may email you occasionally with important news about pyannote models." We stopped there, so we haven't seen WhisperX's labeled output. Its README is direct about what to expect: "Diarization is far from perfect."

In WhisperAI, there's no token and no second website. Leave the number of speakers on Auto or set it from 2 to 10, then rename Speaker 1 to Dana once and it changes on every line. More on transcripts with speaker labels.

Is there a WhisperX online?

Sort of. The README links demos on Replicate for anyone without a GPU, and one of them lists about $0.034 a run, varying with your input. whisperx.app is a separate website that runs the pipeline for you. It calls itself independent, with "no affiliation to the WhisperX authors or OpenAI." It charges $9 a month for 120 minutes or $29 for 600, with paid files up to 60 minutes and 500 MB.

If your recordings run long, that 60-minute limit is the number to look at. WhisperAI takes files up to 10 hours or 5 GB, and Business Pro is $24.99 a month with no minute limit, which beats paying $29 for 600. We don't run WhisperX: AssemblyAI does the transcribing. Keep the page open while the file uploads. After that, close the tab or shut the laptop.

A WhisperX alternative: what changes when you switch

Upload the same file you'd have given WhisperX. There's no model to pick and nothing to keep running. Three things change:

Your transcript gets an editor. Fix a word, then export: SRT or VTT within 30 days of the recording, everything else whenever you like. On any plan, the Claude and ChatGPT connectors can read it too.
Word timings move to the API. Sentence timing in the web app is what subtitles usually need. For anything built on single words, there's the Developer API.
Your audio stays in our cloud until you delete it. The security page has the details.

If you're weighing other ways to run Whisper, see WhisperAI vs OpenAI Whisper or the desktop apps, or our guide to running Whisper itself. To try WhisperAI, upload a short clip on the 5 free minutes.

Questions people also ask

What does WhisperX do?

It's an open-source Python pipeline that transcribes audio with a Whisper model through faster-whisper, lines each word up with the audio using a wav2vec2 model, and can mark who spoke using pyannote.

What is the difference between Whisper and WhisperX?

Whisper is OpenAI's speech recognition model. WhisperX is a separate project built on it: it adds word-level timestamps, speaker labels and batched transcription.

Is WhisperX free?

The code is, under a BSD-2-Clause license. The hardware and the setup are on you.

Is WhisperX made by OpenAI?

No. The WhisperX paper is by Max Bain, Jaesung Huh, Tengda Han and Andrew Zisserman. WhisperAI isn't part of it either: we're an independent product, not affiliated with OpenAI.

Get a transcript with speaker labels without installing anything

Start on the free plan, no card. Upgrade when you need more minutes.

Try WhisperAI free