Skip to main content

Crate voxr_narrate

Crate voxr_narrate 

Source
Expand description

Narration scripts: phrases and chunks a voice speaks, tags, subtitle and timed script files, the written script, numbers read back from words and digits, the check of a spoken take against its script, its glitches, and the polish of its voice. Also a spoken take read into timed words, and the take loop’s cache keys, file names and take-record files, and its pick of each part’s best take through a speech engine and a judge, the designed or prepared voice every take clones, a whole VoxCPM2 run from script to voiceover, the per-phrase run of every other engine, a timed script’s lines placed at their times, the spoken track time-stretched to a pace, and the skill’s WAV files. Also, behind features, the real engines behind the take loop’s seams: Kokoro, VITS and VoxCPM2 as speech engines, Whisper as the hearing, and the runs a caller hands loaded models to. Also the speak and transcribe commands around them: every printed line, file and exit code, and the video manifest’s narration.

The crate implements the speak and transcribe contract of the motion-video skill (https://github.com/farhan-syah/motion-video-skill). “The skill” in these docs means it.

Re-exports§

pub use wav::f32_wav;
pub use wav::pcm16_round_trip;
pub use wav::pcm16_wav;
pub use wav::pcm16_wav_raw;
pub use wav::probed_duration;
pub use wav::read_wav;
pub use wav::read_wav_file;
pub use adapters::*;
pub use check::*;
pub use command::*;
pub use glitch::*;
pub use polish::*;
pub use take::*;
pub use text::*;
pub use transcript::*;

Modules§

adapters
The real engines behind the take loop’s seams: voxr’s Kokoro, VITS and VoxCPM2 models as SpeechEngines, Whisper as the Hearing, and NarrationRuns, the SpeakRuns a caller hands loaded models to.
check
Checks a spoken take against its script: word alignment, the speech check verdict, and script words placed on recognized timings.
command
The speak and transcribe commands around the take loop and the transcriber, as the skill’s speakCmd and transcribeCmd run them: the flags, the script read before any audio, the voice polish and pace after it, the files beside the voiceover, the video manifest’s narration, the words files and the video timeline, every printed line, the exit codes, and the JSON objects --json prints.
glitch
Glitches in a generated take: clicks, edge thumps and cut-off ends, and how far its pitch rises above the voice’s usual pitch.
polish
The voice layer’s polish of a take: sentence levels evened, a gentle compressor, a de-esser and peak normalizing, and the 16-bit samples a polished take is written as. Also a reference voice’s preparation for cloning: breaths turned down and a thin voice’s bass lifted.
take
The take loop’s files, seeds and cache keys: the seed a part’s own text draws, the VoxCPM2 cache key builders and the file names they carry, and the .takes.json/.phrases.json/script text files a take reads and writes. Also a take’s speaking pace: syllables per second and the verdict against a target. Also the seams a take is spoken, stored and judged through, the pick of each part’s best take, and the split of a part that fails whole. Also the voice every take clones: designed from a description or prepared from a user’s recording, and kept in the cache. Also the kept takes shaped and joined into the voiceover, with each phrase’s span and each part’s check, heat and marks. Also the take cache on disk, and a whole VoxCPM2 run from script to voiceover. Also which engine, model, voice and speed a run resolves to speak with, and the per-phrase run of every other engine, the user’s TTS command among them. Also a timed script’s lines placed at their times, and, with the stretch feature, the spoken track time-stretched to a pace. Also a whole speak run from script to voiceover, which picks the branch: one engine run, a time-stretch, or a timed placement.
text
Narration text: phrases and chunks a voice speaks, tags and delivery notes, script files, the written script, and numbers read from words.
transcript
A narration take read into timed words: the take’s loudness, the pieces a recognizer reads, recognizer loops collapsed, word times settled on the sound, script words placed phrase by phrase, and the transcript files. With the whisper feature, Whisper as the recognizer.
wav
Mono WAV files as the skill writes and reads them: 16-bit samples behind a 44-byte header, written with the skill’s own rounding, and read back as ffmpeg decodes them.

Enums§

Error
voxr-narrate errors.

Type Aliases§

Result
voxr-narrate result type.