convez.ai
SkillDados & AnalyticsIntermediário
XqFeng-Josie

VoxKitchen

XqFeng-Josie/VoxKitchen

Pipeline YAML para preparar dados de áudio e fala para ASR/TTS

Ver no GitHub
BomPluginScore · de 0 a 100

Kit declarativo para processamento de dados de fala que transforma gravações brutas em datasets prontos para treinamento via pipelines YAML. Integra 52 operadores para conversão, segmentação, limpeza, diarização e filtragem de áudio, executando tudo em containers Docker isolados.

Instalação

$npx convez-ai XqFeng-Josie/VoxKitchen

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
2
Forks
0
Acompanhando
0
Saúde
71%
Último commit
há 4 meses
Versão
v0.3.1
Issues abertas
0
Contribuidores
1

Casos de uso

  • Preparar datasets de treinamento para modelos ASR (reconhecimento de fala)
  • Processar e inspecionar dados de TTS (síntese de fala)
  • Diarização de falantes, detecção de idioma e métricas de qualidade de áudio
  • Limpeza, filtragem e empacotamento de áudio em escala sem scripts customizados

Veredito

Ferramenta madura e bem-arquitetada para quem precisa processar grandes volumes de dados de áudio de forma reproduzível. Docker-first e resumável por design, elimina scripts frágeis, mas exige familiaridade com pipelines declarativos e stack de áudio.

README

A tradução para o português deste README ainda está na fila.

Use VoxKitchen when you want to:

  • turn long recordings into ASR training data;
  • prepare and inspect TTS datasets;
  • diarize speakers, tag languages, or run speech quality checks;
  • clean, filter, and package audio without maintaining one-off scripts.

Why VoxKitchen

Speech data preparation is usually a chain of fragile scripts: convert audio, split speech, denoise, transcribe, diarize, filter, and export. VoxKitchen makes that chain explicit and repeatable:

  • Docker-first execution: prebuilt runtimes avoid local dependency conflicts.
  • One YAML pipeline: define ingest, stages, filters, and output packs in one file.
  • 52 built-in operators: audio prep, VAD, ASR, diarization, TTS, quality metrics, and packing.
  • Resumable by design: every stage checkpoints under ./work.
  • Inspectable outputs: reports, cut statistics, provenance, and per-stage errors.

Quick Start

Requirements:

  • Docker
  • Python 3.10+ for the lightweight vkit launcher

Install the vkit launcher from PyPI:

pipx install voxkitchen      # recommended — isolates the launcher
# or
pip install voxkitchen

This installs only the lightweight launcher and inspection commands (a few MB, no torch / ASR / TTS dependencies). All pipeline runtime dependencies stay inside the prebuilt Docker images.

Run the included demo with the smallest runtime image. No repository clone is required; the published image includes the demo pipeline and demo audio.

vkit docker pull --tag slim
vkit docker run --tag slim examples/pipelines/demo-no-asr.yaml
vkit inspect run ./work/demo-no-asr
$ vkit docker run --tag slim examples/pipelines/demo-no-asr.yaml
06:20:54  stage [1/9] to_wav  (ffmpeg_convert, 1 cuts in, env=core)
06:20:55  stage [1/9] to_wav  done → 1 cuts out (0.3s)
06:20:55  stage [2/9] vad  (silero_vad, 1 cuts in, env=core)
06:21:08  stage [2/9] vad  done → 7 cuts out (13.2s)
06:21:08  stage [3/9] extract  (ffmpeg_convert, 7 cuts in, env=core)
06:21:08  stage [3/9] extract  done → 7 cuts out (0.6s)
06:21:08  stage [4/9] snr  (snr_estimate, 7 cuts in, env=core)
06:21:08  stage [4/9] snr  done → 7 cuts out (0.0s)
06:21:08  stage [5/9] pitch  (pitch_stats, 7 cuts in, env=core)
06:21:11  stage [5/9] pitch  done → 7 cuts out (2.3s)
06:21:11  stage [6/9] clipping  (clipping_detect, 7 cuts in, env=core)
06:21:11  stage [6/9] clipping  done → 7 cuts out (0.0s)
06:21:11  stage [7/9] gender  (gender_classify, 7 cuts in, env=core)
06:21:44  stage [7/9] gender  done → 7 cuts out (33.2s)
06:21:44  stage [8/9] filter  (quality_score_filter, 7 cuts in, env=core)
06:21:44  stage [8/9] filter  done → 7 cuts out (0.0s)
06:21:44  stage [9/9] pack  (pack_jsonl, 7 cuts in, env=core)
06:21:44  stage [9/9] pack  done → 7 cuts out (0.0s)
06:21:44  report generated: work/demo-no-asr/report.html
pipeline complete
  work_dir: work/demo-no-asr
  final cuts: work/demo-no-asr/08_pack/cuts.jsonl.gz
  report: work/demo-no-asr/report.html

$ vkit inspect run ./work/demo-no-asr
Pipeline run: demo-no-asr
  00_to_wav: OK (1 cuts)  0.3s, 4 cuts/s
  01_vad: OK (7 cuts)  13.2s, 1 cuts/s
  02_extract: OK (7 cuts)  0.6s, 11 cuts/s
  03_snr: OK (7 cuts)  0.0s, 310 cuts/s
  04_pitch: OK (7 cuts)  2.3s, 3 cuts/s
  05_clipping: OK (7 cuts)  0.0s, 381 cuts/s
  06_gender: OK (7 cuts)  33.2s, 0 cuts/s
  07_filter: OK (7 cuts)  0.0s, 8124 cuts/s
  08_pack: OK (7 cuts)  0.0s, 6638 cuts/s

vkit docker run writes run artifacts under ./work and exported datasets under ./output with your host user ID. It also mounts ./data automatically when that directory exists.

What You Can Build

GoalStart withRuntime image
Clean and filter raw speech audiovkit init my-cleaning --template cleaningslim
Build ASR training manifestsvkit init my-asr --template asrasr
Analyze speakers and languagesvkit init my-speakers --template speakerlatest
Prepare TTS training data (quality gate)vkit init my-tts --template ttsasr
Synthesize speech in a built-in voicesee Speaker TTS tutorialtts
Clone a voice from a 3–10 s referencesee Voice Cloning & TTS tutorialtts or fish-speech

How It Works

VoxKitchen pipeline overview

A pipeline is a YAML file. Each stage reads a CutSet, writes a checkpoint, and passes the result to the next stage.

version: "0.1"
name: my-pipeline
work_dir: ./work/${name}-${run_id}

ingest:
  source: dir
  args:
    root: ./data
    recursive: true

stages:
  - name: resample
    op: resample
    args: { target_sr: 16000, target_channels: 1 }

  - name: vad
    op: silero_vad
    args: { threshold: 0.5 }

  - name: asr
    op: faster_whisper_asr
    args: { model: large-v3, compute_type: float16 }

  - name: filter
    op: quality_score_filter
    args:
      conditions: ["duration > 1", "duration < 30", "metrics.snr > 10"]

  - name: pack
    op: pack_jsonl

Interrupted runs resume from completed checkpoints.

Create A Project

vkit init my-project --template asr
cd my-project

# Put your audio files in ./data first.
vkit validate pipeline.yaml
vkit docker run --tag asr pipeline.yaml --dry-run
vkit docker run --tag asr pipeline.yaml
vkit inspect run work/

List templates:

vkit init --list-templates

Not sure which image a pipeline needs? Run:

vkit validate pipeline.yaml

It prints the recommended vkit docker pull --tag ... and vkit docker run --tag ... commands for that YAML.

Runtime Images

Every vkit docker command accepts --tag <name>:

TagUse whenGPUApprox. size
slimCPU-friendly cleaning, VAD, quality, pack, enhancementno~22 GB
asrFaster-Whisper, FunASR, Qwen3-ASR, forced alignmentyes~36 GB
diarizePyannote speaker diarizationyes~20 GB
ttsKokoro, ChatTTS, CosyVoiceyes~46 GB
fish-speechFish-Speech isolated runtimeyes~58 GB
latestMixed pipelines across ASR, diarization, TTS, or Fish-Speechyes~88 GB

Use latest when one pipeline mixes multiple runtime families, such as ASR plus diarization or ASR plus TTS. Otherwise, prefer the smallest image that contains the operators you need.

Useful checks:

vkit docker pull --tag asr
vkit docker doctor --tag asr --expect asr
vkit docker doctor --tag latest

Configuration

Some operators require API tokens. Create ./.env; vkit docker run passes it into the container automatically.

cp .env.example .env
VariableRequired byNotes
HF_TOKENpyannote_diarizeAccept the pyannote model agreement on HuggingFace first.

Common Commands

vkit init <path> --template asr           # Scaffold a project
vkit validate pipeline.yaml               # Validate YAML and recommend an image
vkit docker run --tag asr pipeline.yaml --dry-run
vkit docker run --tag asr pipeline.yaml
vkit inspect run work/                    # Stage summary
vkit inspect cuts <cuts.jsonl.gz>          # CutSet statistics
vkit inspect errors work/                  # Per-stage failed cuts
vkit operators search <keyword>            # Find operators by name or summary
vkit operators --category quality          # List one category's operators
vkit schema export --out pipeline.schema.json  # Editor autocompletion for YAML
vkit recipes                               # List dataset recipes
vkit docker download --tag slim librispeech --root ./data/librispeech --subsets dev-clean
vkit docker doctor --tag latest            # Check image health

Documentation

Agent Skill

The repo includes an agent-neutral VoxKitchen skill at skill/. Claude, Codex, and other SKILL.md-compatible agents can copy, symlink, or import that folder into their own skill search path. The skill follows the Docker-first vkit workflow in this README.

Citation

If you use VoxKitchen in your research, please cite it. The repository ships CITATION.cff, which GitHub renders under "Cite this repository".

License

Apache 2.0. See LICENSE.

Ver o original no GitHub

Bom para

Relacionados