Whisper-TikTok: turning an audio file into a captioned TikTok video with Whisper, Edge TTS and FFMPEG

What Whisper-TikTok actually produces, and for whom

The project takes an audio file and returns a finished vertical video: a transcription from OpenAI's Whisper model, a voiceover synthesized through the Microsoft Edge Cloud text-to-speech API, and a rendered MP4 assembled by FFMPEG. The README describes the intended output as a TikTok video with subtitles and a voiceover, and the repository ships a Streamlit dependency, so the same pipeline is reachable from a browser UI as well as from a terminal.

The audience is narrow and specific. It suits someone who has audio already (a podcast cut, a recording, a downloaded clip) and wants a captioned short without touching a video editor. It does not suit anyone who wants to author a video shot by shot. There is no timeline, no clip trimming, no keyframe control. The tool decides the composition; you supply the audio and a handful of style flags.

The three-stage pipeline: Whisper, Edge TTS, FFMPEG

Stage one is transcription. Whisper-TikTok depends on both openai-whisper and stable-ts, and the acknowledgments credit stable-ts with making the Whisper model usable "in a stable manner with font color and subtitle format options." That pairing matters: stable-ts is what produces word-level timing that survives the subtitle styling step, rather than a plain transcript that would drift against the audio.

Stage two is the voiceover. The edge-tts package (pinned to version 7.2.6 in pyproject.toml) talks to the Microsoft Edge read-aloud service. The README is explicit about why: the author considers that service more natural than the "often monotonous and artificial voiceovers prevalent in numerous TikTok videos." This is a cloud call, not a local model, so the voice quality depends on a service the project does not control.

Stage three is rendering. FFMPEG composites the background video, the subtitle track and the audio into the final file. The README states plainly that FFMPEG binaries "are not included in the repository and must be installed separately," and that they must be on your system PATH. Everything upstream can succeed and the render still fail if that binary is missing.

The dependency list also includes yt-dlp and tiktok-uploader, which is consistent with the README example that accepts a YouTube URL as a background video source.

Installing Whisper-TikTok and rendering a first video

The README states the project has been tested on Windows 10, Windows 11 and Ubuntu 24.04 with Python 3.11 and 3.12, and pyproject.toml enforces requires-python >=3.11. Clone the repository first:

git clone https://github.com/MatteoFasulo/Whisper-TikTok.git

The README documents two dependency paths. The plain pip route uses requirements.txt, which is generated by uv and pins every transitive package with hashes:

pip install -r requirements.txt

The README says the project encourages uv instead, because the torch backend can be selected automatically for your hardware:

uv pip install -e . --torch-backend=auto

FFMPEG is a separate install and is not vendored. The README lists the package-manager commands per platform; on Ubuntu or Debian it is:

sudo apt update && sudo apt install ffmpeg

Once the console script is on your PATH, the README's entry point is the help command, which lists the available subcommands:

whisper_tiktok --help

The create subcommand carries the styling flags, and the README points you at it for the full list:

whisper_tiktok create --help

A first real run can pick a named voice, which is the README's first usage example:

whisper_tiktok create --tts en-US-EricNeural

If you would rather not choose a voice yourself, the README documents a random option that requires both a gender and a language, and notes that Whisper auto-detects the language of the audio and uses the corresponding model:

whisper_tiktok create --random_voice --gender Male --language en-US

To see what voices exist before picking one, the README provides a listing command that reports tone, style and suitable scenarios for each entry:

whisper_tiktok list-voices

Where the tool stops being the right choice

The hard external dependency is FFMPEG. If you cannot install binaries on the machine (a locked-down container, a managed laptop without package-manager rights), the project cannot render anything, regardless of how well the Whisper and TTS stages behave. The README does not document a bundled or pure-Python fallback for the render step.

The second constraint is the cloud TTS call. Because the voiceover comes from the Edge read-aloud service through edge-tts, any network restriction, service change or rate limit affects output. The README does not document an offline voice path, and it does not describe retry or fallback behavior for that call.

The third is scope. Whisper-TikTok is a generator, not an editor. If your task is cutting a two-minute clip out of a longer recording, or matching captions to a specific on-screen moment, the CLI surface shown in the README (voice, subtitle format, subtitle position, font size, font color) is styling, not timing control. For that kind of work a general-purpose editor or a scripted FFMPEG pipeline you write yourself gives you the control this tool deliberately abstracts away.

How it differs from plain Whisper or a hand-written FFMPEG script

The obvious alternative is running openai-whisper directly and then assembling the video yourself. Whisper alone gives you a transcript and timestamps; it does not give you a voiceover, subtitle styling, or a vertical render. Whisper-TikTok's contribution is the glue: stable-ts for timing that holds up under styling, edge-tts for the voice, and FFMPEG for composition, all behind one command.

The second alternative is writing the FFMPEG invocation yourself. That is genuinely more flexible. You control the filter graph, the subtitle burn-in, the audio mixing and the encoding parameters. The trade-off is that you also maintain all of it, and you still have to solve transcription and text-to-speech separately. Whisper-TikTok's value proposition is that it collapses three tools into one CLI with a documented flag set, at the cost of the control you would have had over the render stage.

A third comparison worth noting is the Streamlit dependency. The project ships both a CLI and a browser interface over the same pipeline, which is unusual for a tool of this size and suggests the author wanted non-terminal users to reach it without a separate front end.

Licence, maintenance and the cost of upgrading

Whisper-TikTok is licensed under Apache-2.0, as stated in the README and in pyproject.toml. That is a permissive licence with an explicit patent grant, which is generally the easier end of the spectrum for commercial use, but the obligations (notice retention, stating changes) still apply and depend on how you redistribute. Nothing here is legal advice; read the LICENSE file in the repository.

The dependency surface is the real upgrade cost. edge-tts is pinned exactly (==7.2.6), while openai-whisper, stable-ts, torch, streamlit and others use lower bounds. An exact pin on the TTS client means voice behavior is stable but a fix upstream will not reach you without a version bump. The torch floor (>=2.6.0) is the heaviest install in the list, and the README's uv path exists specifically because picking the right torch backend by hand is error-prone.

The last push to the repository was on 2026-07-10, and the most recent release listed is v2.0.4 from 2026-01-25. The repository is not archived. The project ships a CHANGELOG.md, a pre-commit configuration and a pytest suite under tests/, so there is a maintenance structure in place, but the README does not document a deprecation or migration policy for breaking CLI changes.

Editorial conclusion

Adopt Whisper-TikTok if you already have an audio file and want a captioned vertical video without opening an editor, and you are comfortable with Python 3.11 or 3.12 and a separate FFMPEG install. Skip it if you need frame-level creative control, a hosted service, or a pipeline that runs without FFMPEG present. Before committing, run whisper_tiktok create --help on your own machine and confirm that the subtitle position, font and voice flags you need are actually exposed, and check that your torch install matches the backend uv picks for your hardware.

Frequently asked questions

Official sources

1. License: Apache-2.0
2. MatteoFasulo/Whisper-TikTok on GitHub
3. Project website
4. README
5. Releases