Data Preparation

Create Speech Dataset

Drop audio + transcript files here
or click to browse (individual files or .zip)

Supported: .wav, .mp3, .m4a, .sph + .srt, .vtt, .txt (or .zip)
Max: 10GB per file, 10GB total (up to 10 ZIPs per session)
For more than 10 audio files, upload a ZIP archive.

Matching: Files are paired by name (e.g., audio1.wav ↔ audio1.vtt)
Note: .txt files (no timestamps) require audio under 5 minutes. Use .srt/.vtt for longer audio.

How to create a ZIP ▾

Place your audio + transcript files together and zip them:

Mac: Select all files → right-click → Compress
Windows: Select all files → right-click → Send to → Compressed (zipped) folder
Linux/CLI: zip dataset.zip *.wav *.vtt

How it works ▾

Upload paired audio + transcript files (.srt/.vtt/.txt)
Process: forced alignment + smart chunking (≤30s)
Output: HuggingFace dataset ready for training

HuggingFace Token

Required to push dataset to Hub. Get token

Organization / User

Dataset Name

Make dataset private

Language

Language of the audio/transcripts (ISO 639-3). Affects text alignment accuracy.

Dataset Split Options

Train split only (no validation)

Create validation split (10% or max 50 rows)

Push as validation split only (no train split)

Advanced Audio Settings ▾

VAD Threshold: 0.5

Lower = more sensitive (detects more speech). Default: 0.5

Min Silence Duration (ms)

Silence gap to trigger a segment split (speech padding = half this value)

Target Chunk Duration (s)

Soft target for chunk length. Ignored if Pack is off. Default: 20s

Max Chunk Duration (s)

Hard cap (max 30s for Whisper). Default: 30s

Min Chunk Duration (s)

Filter out chunks shorter than this (always applied). Default: 5s

Pack speech regions (remove silence)

When enabled, speech regions are concatenated and silence is removed. When disabled, each VAD segment becomes a separate chunk with original timing.

Audio Cleaning

None (default)

Basic - Highpass 80Hz + normalize volume

Denoise - Basic + spectral noise reduction

Applied to speech chunks after VAD extraction. Basic removes low-frequency rumble and normalizes volume. Denoise additionally reduces background noise (fans, AC, etc).

Enable quality filtering

Filter out poorly-aligned chunks based on character density (keeps 6–32 chars/sec, typical speech is ~20–25 chars/sec). Min chunk duration always applies separately.
Note: transcripts with digits (e.g. "1472" vs "one four seven two") have fewer characters for the same spoken duration, which may cause false filtering. Consider disabling for number-heavy data.

Current Job

No job

No processing running. Upload files and click Process to start.

Processing History

Time	Files	Samples	Audio Prepared	Cost	Output	Status
No data preparation jobs yet