voxr

Audio and speech on top of boostr models and numr tensors.

boostr holds the model networks and their tensor kernels. voxr holds everything around them that works on samples, files, text and voices.

What it holds

Always built:

wav: RIFF/WAVE read and write, and stream headers of unknown length.
resample: a general polyphase resampler, plus resamplers that match ffmpeg's default output, SciPy's resample_poly and librosa's default output.
stretch: a WSOLA tempo engine that matches ffmpeg's atempo output.
enhance: the reference-take chain of denoise, filters, loudness and limiter.
metrics and sound_review: single-sound measures and the sound review of the motion-video skill (https://github.com/farhan-syah/motion-video-skill).
pitch, quality, spectral: pitch tracking, take quality and two-signal spectral distance.
eval: word and character error rates for ASR.
tts_eval: the signal and summary halves of the TTS quality gate.

Behind features:

Feature	Adds
`decode`	Compressed-audio decode through symphonia, with ffmpeg-matched trims and downmix
`spacy-tokenizer`	The spaCy `en_core_web_sm` 3.8.0 tokenizer
`spacy-tagger`	The `en_core_web_sm` tagger
`misaki`	misaki English G2P with embedded US and GB lexicons
`tts`	`TtsEngine`, `TtsBundle` and `Voice`
`kokoro`	Kokoro synthesis
`vits`	VITS (MMS-TTS) synthesis
`voxcpm`	VoxCPM2 voice cloning
`whisper`	Whisper transcription with word timings
`vad`	Silero VAD segmentation
`clap`	CLAP audio-text scoring
`neucodec`	The NeuCodec model gate
`corpus`	Speech-corpus preparation
`f16`	Half-precision tensors
`cuda`	The CUDA backend. Needs CUDA 12.8 or newer, or 13.x

The repository README lists what each feature implies.

No weights are compiled in. Each engine loads its checkpoint at run time from a directory, and voxr fetches nothing over the network.

Apache-2.0. Ported code and redistributed data are listed with their licences in LICENSE-OTHERS.