AI Voice Studio 🎧 — phone voice note ➜ studio mic

Turn your phone voice note into a studio-mic voice. Free, local, one command.

Same take, same phone. Turn the sound on 🔊

https://github.com/user-attachments/assets/ad3ca9ae-6334-4b67-8e34-ae149464f1f7

No noise, no room echo, the treble lost to phone compression rebuilt, long pauses tightened, mixed to the loudness Instagram and TikTok expect. Made by Amine (@justbyamine): every Darija tech reel on the account is voiced this way.

Quick start

You have	Do this
No GPU, nothing installed	Click Open in Colab above (or import ai_voice_studio.ipynb in Kaggle), Run all, open the link it prints
A PC (Windows/Linux)	Install (below), then python app.py and drop your voice note in the browser
A terminal habit	python voice_pro.py voice-note.mp4 output

Demo

Before / after audio only: demo/before-after.mp3 (video above: demo/before-after.mp4, made with tools/make_demo_video.py so you can build the same comparison for your own voice).

The reel about this tool, as published on TikTok (downloaded from TikTok Studio, with the @justbyamine watermark). It plays the raw phone voice, then the same voice after the script; turn the sound on:

https://github.com/user-attachments/assets/dbd93934-ddee-4c4d-9f98-8867de06e567

Watch it on TikTok: https://www.tiktok.com/@justbyamine/video/7698607101580496180 (File in the repo: demo/voice-pro-reel-from-tiktok.mp4.)

Proof: used on my published reels

• Reel about this tool, Instagram: https://www.instagram.com/reel/DetAeRqA1k8/
• Same reel on TikTok: https://www.tiktok.com/@justbyamine/video/7698607101580496180
• Same reel on Facebook: https://web.facebook.com/reel/1664386081973860/

Follow: Instagram @justbyamine · TikTok @justbyamine · Facebook: Just by amine · GitHub @OutOfAmine

What it does

Step	Output	What happens
1. AI restoration	1-restored.wav	resemble-enhance removes noise and room echo and rebuilds the treble lost to phone compression
2. Pause tightening	2-tight.wav	pauses longer than 180 ms are shortened to 140 ms (8 ms crossfades)
3. Broadcast chain	3-chain.wav	warmth (+3 dB @120 Hz), less "phone box" (−3 dB @380 Hz), presence (+2.5 dB @4.8 kHz), air, de-esser, 2-stage compression
4. Final	Voice-Pro-Final.wav	quiet room tone at −66 dBFS (no dead digital silence), gain kept in float, peak limiter, then one measured linear gain to −14 LUFS with true peak ≤ −1.5 dBTP (Instagram / TikTok level, no pumping)

It also writes time-map.json (old → new timings) so subtitles made on the original voice can be re-synced. No pitch or speed change.

Video in, video out: give it a video (.mp4, .mov, .webm) and it also writes Voice-Pro-Video.mp4: the same picture (not re-encoded) with the new voice. Pauses are kept for videos so the voice stays on the lips (VOICE_TIGHTEN=1 forces the tightening anyway; VOICE_TIGHTEN=0 turns it off for audio too).

Install (Windows, Python 3.10)

python -m venv .venv
.venv\Scripts\python -m pip install torch torchaudio --index-url https://download.pytorch.org/whl/cu121
.venv\Scripts\python -m pip install --no-deps resemble-enhance
.venv\Scripts\python -m pip install -r requirements.txt

No NVIDIA GPU? Install the normal torch (pip install torch torchaudio); it works on CPU, just slower. The script stubs deepspeed (only needed for training) so resemble-enhance imports on Windows.

Use

Web UI (upload or record, pick Strong / Soft, download the result):

On Windows, double-click Launch Studio.cmd. The app opens in your browser. Keep the terminal open while using the studio; closing it stops the app.

• 1. Add your voice: choose Upload a file or Record now, add your audio or video, then click Continue.
• 2. Clean up: click Make my voice better. The recommended settings are already selected. Naming the recording, choosing Strong / Soft and turning off Shorten long pauses are under Optional settings.
• 3. Download: play Before and After side by side, then click Download my voice. A video input also shows the video with the new voice, ready to download from its player.
• Saved voices keeps completed recordings, originals and processing results between app restarts. Files are stored in local-library/ beside app.py and excluded from Git. Set VOICE_STUDIO_LIBRARY to choose another folder.

The default app listens only on 127.0.0.1. Microphone access requires browser permission; if the embedded browser has no microphone, open the local URL in Chrome or Edge. The first enhancement may download model weights; subsequent processing uses local files. --share remains an explicit opt-in for Colab: it exposes the app and library through a public link.

.venv\Scripts\python app.py

Command line:

.venv\Scripts\python voice_pro.py voice-note.mp4 output

Any audio or video input works (.mp4, .opus, .ogg, .m4a, .mp3, .wav). The result is output/Voice-Pro-Final.wav. The first run downloads the models.

Tips

• Record close to the phone, in a small room with soft things around (less echo).
• Cut your mistakes before running it; the script only shortens silences.
• Sounds "over-processed"? Run with a softer restoration: set VOICE_LAMBD=0.5 (PowerShell: $env:VOICE_LAMBD="0.5").
• Sounds hissy? Turn the 11 kHz air shelf down (it boosts highs the AI rebuilt): set VOICE_AIR_DB=1 (default 3, 0 = off).
• Very soft syllables at the end of a phrase can be erased by the restoration: listen to the result, and re-run with VOICE_LAMBD=0.5 if a word disappears.

---

Darija

Script li kaydir l voice note dyal tilifoun soot b7al micro pro: kay7yed l'echo w l'souffle, kay9sser les silences twal, kayzid chaleur w clarté, w kayjib l'volume dyal Instagram/TikTok. Kolchi f PC dyalek, fabor.

Credits and license

resemble-enhance (MIT) and ffmpeg. This repository: MIT, see LICENSE.