Speech acceleration that doesn't butcher consonants. Speeds up audiobooks and podcasts at 2x-4x while keeping the words intelligible.

Most time-stretching tools (sox, Rubber Band) apply a uniform rate across the whole signal. That's fine up to about 1.5x, but past 2x the consonants get mushy and everything sounds like it's underwater. Osmium allocates its time budget the way a natural fast speaker does: it protects consonant transients and compresses the parts you don't need (sustained vowels, pauses, breaths).

Requires Python 3.11+ and ffmpeg.

uv tool install '.[neural,demucs]'

This installs osmium as a command with all the extras (Mimi neural importance, Demucs source separation). On Apple Silicon, add MLX for GPU acceleration:

uv tool install --with mlx '.[neural,demucs]'

For development:

git clone https://github.com/user/osmium
cd osmium
uv sync

osmium audiobook.mp3 -s 3.0 -o output.mp3
osmium podcast.m4a -s 2.5 -o output.m4a
osmium chapter.wav -s 2.0 --uniform -o output.wav   # skip importance analysis
osmium noisy.mp3 -s 3.0 --denoise deep -o clean.mp3 # adaptive denoising
osmium noisy.mp3 -s 3.0 --denoise none -o raw.mp3   # disable denoising

Stream to speakers:

osmium input.mp3 -s 3.0 --stream | ffplay -nodisp -f f32le -ar 24000 -ac 1 -

Export the importance map without processing:

osmium input.mp3 -s 3.0 --analyze-only -o importance.json

1. Denoise -- spectral gating removes background hiss (on by default; --denoise deep for adaptive mode, --denoise demucs for full source separation)
2. Analyze -- compute per-frame importance from the mel spectrogram (spectral flux + energy, with a 2.5x boost for high-frequency consonant bands)
3. Schedule -- convert importance to a variable rate curve with contrast compression (--rate-gamma), hitting the target speed while giving more time to important frames
4. Stretch -- resample the mel spectrogram according to the rate curve with adaptive smoothing (more smoothing where compression is high, less where consonants need sharp attacks), then reconstruct audio with the Vocos neural vocoder
5. Post-process -- three-stage cleanup of vocoder output: (a) declick removes transient energy spikes from ISTFT phase discontinuities, (b) subtle room ambience adds early reflections that perceptually mask remaining artifacts, (c) warm dither adds low-frequency shaped noise for natural warmth. All on by default, individually toggleable.

See ARCHITECTURE.md for the full picture.

samples/clips/ contains short clips (15s and 30s) extracted from the public domain LibriVox recording of Moby Dick by Herman Melville. These are used for evaluation and listening comparisons.

Generate accelerated versions across all speeds and modes:

scripts/generate_accelerated.sh

This produces MP3s in samples/clips/accelerated/{speed}/{mode}/ for speeds 2x–3.8x and modes:

• uniform — flat rate, no importance analysis (--uniform)
• no-mimi — mel-based importance (default)
• neural — Mimi neural codec importance (--mimi)
• gate-denoise — spectral gating + mel importance
• deep-denoise — adaptive denoising + mel importance
• demucs-denoise — Demucs source separation + mel importance

uv run scripts/eval_wer.py samples/clips/*.wav -s 3.0                           # Whisper WER
uv run scripts/eval_wer.py samples/clips/*.wav -s 3.0 --sweep rate_gamma 1.0 1.5 2.0  # sweep gamma
uv run scripts/abx_test.py version_a.wav version_b.wav                          # ABX listening test

MIT