Examples & Use Cases

Use this page when you already know what kind of speech data task you want to run. Start with a template for normal projects; use the bundled example pipelines for quick checks, demos, and advanced operator combinations.

Quick Demo

The published Docker images include demo pipelines and demo audio, so you can try VoxKitchen without cloning the repository:

vkit docker run --tag slim examples/pipelines/demo-no-asr.yaml --dry-run
vkit docker run --tag slim examples/pipelines/demo-no-asr.yaml
vkit inspect run ./work/demo-no-asr

Use this path to check that Docker, mounts, checkpoints, reports, and inspect commands work on your machine.

Start From A Template

Templates are the recommended starting point for real projects because they create a local project directory with data/, pipeline.yaml, and a short README.

Goal	Command	Runtime image
Clean and filter raw speech audio	vkit init my-cleaning --template cleaning	slim
Build ASR training data	vkit init my-asr --template asr	asr
Analyze speakers and languages	vkit init my-speakers --template speaker	latest
Prepare TTS training data	vkit init my-tts --template tts	asr

Typical run:

cd my-asr
cp /path/to/audio/* data/
vkit docker run --tag asr pipeline.yaml --dry-run
vkit docker run --tag asr pipeline.yaml
vkit inspect run work/

Bundled Example Pipelines

These YAML files are available inside the published Docker images under examples/pipelines/. Clone the repository only if you want to inspect or edit the files locally.

Pipeline	Use case	Runtime image
minimal.yaml	Identity passthrough to check the runner	slim
demo-no-asr.yaml	Small CPU-friendly demo with bundled audio	slim
demo-full.yaml	Full demo with VAD, quality metrics, ASR, gender, filtering	asr
dir-resample-pack.yaml	Directory ingest, resample, normalize, Kaldi export	slim
data-cleaning.yaml	Quality metrics, dedup, filtering, JSONL export	slim
asr-training-data.yaml	VAD, augmentation, ASR labeling, HuggingFace export	asr
librispeech-asr.yaml	Recipe ingest from LibriSpeech, ASR, quality filter	asr
qwen3-asr.yaml	Qwen3-ASR transcription path	asr
forced-align.yaml	Word alignment for existing text/audio	asr
speaker-analysis.yaml	VAD, diarization, speaker/language annotations	latest
speaker-embed.yaml	Extract speaker embeddings from speech segments	slim
tts-data-prep.yaml	Clean, segment, transcribe, align, and pack TTS data	asr
tts-synthesis.yaml	Run a TTS synthesis operator	tts
tts-speaker-filter.yaml	Filter or inspect data by speaker metadata	slim
speech-enhance.yaml	Speech enhancement / denoising	slim
augmentation.yaml	Basic data augmentation flow	slim
noise-augment.yaml	Add noise augmentation from a noise directory	slim
reverb-augment.yaml	Add reverberation augmentation	slim
emotion-recognize.yaml	Emotion annotation	asr
codec-tokenize.yaml	Audio codec token extraction	slim
fleurs-multilingual.yaml	Multilingual recipe-style processing	asr

Validate before running when you are unsure about paths, arguments, or image choice:

vkit docker run --tag asr examples/pipelines/asr-training-data.yaml --dry-run

To edit these examples or write your own pipeline from scratch, see the Pipeline YAML reference.

Choosing The Right Image

Use the smallest image that contains the operators in your pipeline:

Image	Choose it for
slim	CPU-friendly prep, VAD, enhancement, quality checks, codec tokenization, packing
asr	ASR, forced alignment, and emotion annotation
diarize	Pyannote diarization only
tts	TTS engines except Fish-Speech
fish-speech	Fish-Speech isolated runtime
latest	Mixed pipelines across ASR, diarization, TTS, and Fish-Speech

vkit validate pipeline.yaml prints the recommended pull/run command for your pipeline.

Inspect Outputs

Most examples write checkpoints and reports under ./work; exporters usually write final datasets under ./output.

vkit inspect run work/
vkit inspect cuts work/<run>/<stage>/cuts.jsonl.gz
vkit inspect errors work/

For pack_huggingface, the audio column is embedded in the Arrow dataset. If your training code needs decoded arrays, install torchcodec; for metadata or custom decoding, read with datasets.Audio(decode=False).