CrispVoice

Studio-quality voice enhancement that never uploads your voice.

Live demo with before/after audio: crispvoice.francium.tech

Take any voice recording - a phone memo, a laptop-mic capture, a noisy call - and get back clean, full-bodied, podcast-ready audio. Runs 100% locally: your audio never leaves your machine.

Inspired by the excellent cleanvoice.ai and tools like Adobe Podcast Enhance. They showed what is possible; this project open sources how it is done, for anyone who wants the same result without sending their voice to a server.

What it does

"Studio quality" is two problems, and this pipeline solves both:

1. Clarity front-end - a speech-enhancement transformer (ClearerVoice MossFormer2) pre-cleans the recording. Feeding the re-synthesis a clean input measurably improves word intelligibility of the final output.
2. Restoration - a generative model (Resemble Enhance, MIT) re-synthesizes the voice at 44.1 kHz, adding back the bandwidth and body a cheap microphone never captured.
3. Texture blend - 25% of a DeepFilterNet-denoised copy is mixed back in. Re-synthesis alone sounds subtly synthetic; the blend restores natural transients, breaths, and room decay. This stage came out of blind A/B testing against a commercial reference and moved the quality from "processed" to "studio."
4. Mastering - a broadcast-style ffmpeg chain: chest warmth, articulation and air EQ, de-essing, gentle 2:1 compression, and EBU R128 loudness normalization to -19 LUFS.

See ARCHITECTURE.md for the full technical breakdown, including how the mastering EQ was fitted band-by-band against a commercial tool's output.

Hear it

Three short clips, before and after (or use the players on the demo page):

Voice	Before	After
Interview, room noise	demo/noise.m4a	demo/crispvoice.mp3
Broadcast clip, compressed	demo/woman.m4a	demo/woman_crispvoice.mp3
Casual memo, mild background noise	demo/umm.m4a	demo/umm_crispvoice.mp3

Quickstart

Requires uv and curl. Tested on macOS (Apple Silicon); Linux x86_64 should work. ~9 GB disk total (three model runtimes; every byte lives inside this folder and uninstalls with it).

git clone https://github.com/Francium-Tech/CrispVoice.git
cd CrispVoice
./setup.sh                    # one-time: local Python + PyTorch + model weights
./enhance recording.mp3       # writes recording_studio.mp3

More options:

./enhance in.mp3 out.wav          # output format follows the extension; use .wav for lossless
./enhance in.mp3 --preview 30     # quick sample: process only the first 30 seconds
./enhance in.mp3 --nfe 32         # ~2x faster, slightly lower quality (default 64)
./enhance in.mp3 --blend 0.4      # more natural texture (default 0.25)
./enhance in.mp3 --preset natural # lighter compression, more dynamics
./enhance in.mp3 --device cpu     # skip the Apple GPU fast path (auto by default)
./enhance in.mp3 --no-clarity     # skip the ClearerVoice pre-clean stage
./enhance in.mp3 --denoise-only   # cleanup without generative re-synthesis
./enhance in.mp3 --threads 2      # use even less CPU
./enhance in.mp3 --no-master      # skip the mastering chain

Progress, ETA, and memory usage are printed for every chunk.

Design principles

• Private by default. No network calls after setup. Ever.
• Self-contained. The Python interpreter, PyTorch, model weights, and even the ffmpeg binary all live inside the project folder. Nothing is installed system-wide. ./uninstall.sh removes every trace in one go.
• Never hog the machine. CPU-only (deliberately: on Apple Silicon the GPU shares unified memory with the OS, and PyTorch's MPS backend can exhaust it and freeze the machine - a lesson learned the hard way). Half the cores, low process priority, bounded memory via 10-second chunks. On Apple Silicon the heavy model stage runs on the GPU by default - in an isolated, memory-capped subprocess that falls back to CPU on any failure - bringing an 8-minute recording to roughly 5 minutes. Pure CPU (--device cpu) is roughly 2.5x real time (~20 minutes for the same file).

Credits

• Resemble AI for the MIT-licensed enhancement model that does the heavy lifting.
• cleanvoice.ai for the inspiration and the quality bar to chase.
• FFmpeg for decoding, mastering, and encoding.

About

CrispVoice is an open source project by Francium Tech.

License

MIT. Model weights are downloaded separately from ResembleAI/resemble-enhance (also MIT) and are not part of this repository.