Crate voxr 

Sample-level audio DSP, file codecs, voices, and product pipelines.

Placement rule: code that operates on Tensor<R> with backend kernels belongs in boostr::model::audio. Code that operates on Vec<f32> samples, files, directories, voices, or text pipelines belongs here.

Always compiled: RIFF/WAVE in and out (wav), the polyphase resamplers (general, ffmpeg-exact and SciPy-exact), the single-sound metrics and the sound_review built on them, the reference-take enhance chain, pitch, quality, two-signal spectral distance, the atempo-compatible tempo stretch, ASR eval metrics, the signal and summary halves of the tts_eval quality gate, and the Lang tag a voice carries. Feature gates add the compressed-audio decoders, the clap audio-text scorer, the misaki G2P backend (g2p), the spaCy en_core_web_sm tokenizer and tagger, the tts surface, the kokoro and voxcpm engines, the whisper bundle, the vad segmenter, corpus preparation, and the pace, intelligibility and timbre legs of tts_eval.

sound_review implements the sound review of the motion-video skill (https://github.com/farhan-syah/motion-video-skill). Docs here call it “the motion-video skill” or “the skill”.

Re-exports

pub use error::Error;

pub use error::Result;

pub use corpus::CorpusOptions;

pub use corpus::MAX_UTTERANCE_SECS;

pub use corpus::PRETRAINED_TOKENIZER_NAMES;

pub use corpus::SpeechCorpusBuilder;

pub use corpus::TextTokenizer;

pub use corpus::Utterance;

pub use corpus::check_max_speech_duration;

pub use corpus::pack_utterances;

pub use corpus::pack_utterances_with_layout;

pub use decode::decode_audio;

pub use decode::decode_audio_file_mono_at;

pub use decode::decode_audio_like_ffmpeg;

pub use decode::decode_audio_mono_at;

pub use decode::extension_hint;

pub use eval::ErrorRate;

pub use eval::align;

pub use eval::align;

pub use eval::by_group;

pub use eval::canonical_character_error_rate;

pub use eval::canonical_word_error_rate;

pub use eval::canonicalize;

pub use eval::character_error_rate;

pub use eval::grand_total;

pub use eval::normalize;

pub use eval::normalize;

pub use eval::total;

pub use eval::word_error_rate;

pub use ffmpeg::ffmpeg_program;

pub use ffmpeg::ffprobe_program;

pub use lang::Lang;

pub use lang::LangTag;

pub use pitch::PitchOptions;

pub use pitch::PitchTrack;

pub use pitch::estimate_pitch;

pub use quality::TakeQuality;

pub use quality::measure_quality;

pub use resample::DEFAULT_TAPS_PER_PHASE;

pub use resample::MAX_FILTER_TAPS;

pub use resample::PolySample;

pub use resample::ffmpeg_default_layout;

pub use resample::ffmpeg_resampled_len;

pub use resample::librosa_output_len;

pub use resample::resample;

pub use resample::resample_interleaved_like_ffmpeg;

pub use resample::resample_like_ffmpeg;

pub use resample::resample_like_librosa;

pub use resample::resample_poly;

pub use resample::resample_with_taps;

pub use resample::to_mono_at_rate;

pub use resample::to_mono_like_ffmpeg;

pub use spectral::SpectralDistance;

pub use spectral::multi_resolution_stft_distance;

pub use spectral::snr_db;

pub use stretch::MAX_TEMPO;

pub use stretch::MIN_TEMPO;

pub use stretch::PcmSample;

pub use stretch::atempo;

pub use stretch::atempo_interleaved;

pub use stretch::atempo_pcm16;

pub use tts::KOKORO_VOICES;

pub use tts::SynthesizeOptions;

pub use tts::TtsBundle;

pub use tts::TtsEngine;

pub use tts::TtsError;

pub use tts::Voice;

pub use tts::default_kokoro_voices;

pub use tts_eval::F0Stats;

pub use tts_eval::TimbreProxy;

pub use tts_eval::TimbreScorer;

pub use tts_eval::timbre_proxy;

pub use tts_eval::Intelligibility;

pub use tts_eval::IntelligibilityClient;

pub use tts_eval::IntelligibilityScorer;

pub use tts_eval::MetricStat;

pub use tts_eval::RowScore;

pub use tts_eval::SignalStats;

pub use tts_eval::Summary;

pub use tts_eval::Thresholds;

pub use tts_eval::UNLABELLED_AXIS;

pub use tts_eval::signal_stats;

pub use tts_eval::violations;

pub use tts_eval::Pace;

pub use tts_eval::pace;

pub use tts_eval::pace;

pub use tts_eval::pace_from_segments;

pub use vad::SpeechSegment;

pub use vad::VadSegmentOptions;

pub use vad::segments_from_probabilities;

pub use vad::speech_timestamps;

pub use wav::WavData;

pub use wav::WavF64;

pub use wav::decode_wav;

pub use wav::decode_wav_f64;

pub use wav::encode_f32_raw;

pub use wav::encode_pcm16_raw;

pub use wav::encode_wav_f32;

pub use wav::encode_wav_pcm16;

pub use wav::encode_wav_pcm16_values;

pub use wav::to_mono;

pub use wav::wav_stream_header_f32;

pub use wav::wav_stream_header_pcm16;

pub use whisper::TranscribeOptions;

pub use whisper::Transcription;

pub use whisper::WhisperBundle;

pub use whisper::WhisperGenerationConfig;

pub use whisper::WordTiming;

Modules

clap

CLAP scoring of audio against text: a checkpoint directory loaded into boostr’s ClapModel, its tokenizer, and the cosine scores between clips and texts.

corpus

Speech-corpus preparation: a decoded recording in, a packed speech-LM token stream out.

decode

Compressed-audio decoding via symphonia: WAV, FLAC, MP3, OGG/Vorbis, AAC and ALAC in MP4/M4A/MOV, and Opus or Vorbis in WebM/MKV/OGG, video files’ audio tracks included. Each file decodes as ffmpeg decodes it: encoder priming and padding trimmed.

enhance

Turn a raw recording into a usable reference.

error

voxr error types.

eval

Error-rate metrics for ASR output and for the intelligibility leg of TTS eval.

ffmpeg

The ffmpeg programs voxr’s ffmpeg-matching code is checked against: $FFMPEG and $FFPROBE, else ffmpeg and ffprobe on PATH.

g2p

Grapheme-to-phoneme (G2P) conversion.

kokoro

Kokoro synthesis pipeline: G2P, voice-pack lookup on disk, and the KokoroEngine that drives boostr’s KokoroModelV2.

lang

Lang: the language tag a crate::tts::Voice carries.

metrics

Single-signal measures of a finished sound, each a port of the NumPy and SciPy computation the motion-video skill’s sound review runs: sharpness, the share of energy above 8 kHz, the tail against the loudest 50 ms, and the 4x oversampled true peak.

pitch

Fundamental-frequency (F0) estimation with YIN.

quality

Take-quality metrics for recorded/decoded speech: peak, RMS, noise floor, SNR.

resample

Sample-rate conversion for f32 audio.

sound_review

The motion-video skill’s developer sound review (sound-review.py): numeric limits on every rendered sound, and recognizability of the base variants against their says texts by an audio-text model.

spectral

Multi-resolution spectral distance and two-signal SNR, for comparing a decoded/rendered take against its reference (e.g. an AudioVAE decoded at F16 vs the same latent decoded at F32).

stretch

Tempo change without a pitch change, by waveform-similarity overlap-add.

tts

Engine-agnostic text-to-speech surface: the TtsEngine trait a model implements and the TtsBundle that owns one plus its voice catalog.

tts_eval

TTS quality gate: numbers for every rendered clip.

vad

Segmentation: speech probabilities in, utterance boundaries out.

voxcpm

VoxCPM2 voice-cloning pipeline: reference recordings decoded from a voices directory or encoded per request, text tokenized through splintr, and the VoxCpm2Engine that drives boostr’s VoxCpm2Model. Weights-source selection (VoxCpm2Weights) stays with the model in boostr.

wav

RIFF/WAVE: the one container this crate reads and writes without the decode feature. decode parses PCM and IEEE-float files through hound, decode_f64 reads them to f64 as libsndfile does, encode writes them, stream writes the unknown-length header a streamed response starts with, format holds the format tags and canonical header layout all sides agree on, and layout parses the fmt chunk and the channel layout it names.

whisper

Whisper as a product pipeline: the checkpoint’s weights from boostr plus the tokenizer and decoding constraints that make it transcribe.