| --- |
| tags: |
| - audio |
| - speech |
| - whisper |
| - dataset |
| --- |
| # sample-dataset-test-neural-pack |
|
|
| Speech dataset prepared with Trelis Studio. |
|
|
| ## Statistics |
|
|
| | Metric | Value | |
| |--------|-------| |
| | Source files | 6 | |
| | Train samples | 136 | |
| | Validation samples | 15 | |
| | Total duration | 62.6 minutes | |
|
|
| ## Columns |
|
|
| | Column | Type | Description | |
| |--------|------|-------------| |
| | `audio` | Audio | Audio segment (16kHz) - speech only, silence stripped via VAD | |
| | `text` | string | Plain transcription (no timestamps) - backwards compatible | |
| | `text_ts` | string | Transcription WITH Whisper timestamp tokens (e.g., `<|0.00|>Hello<|0.50|>`) | |
| | `start_time` | string | Segment start in original audio (HH:MM:SS.mmm) | |
| | `end_time` | string | Segment end in original audio (HH:MM:SS.mmm) | |
| | `speech_duration` | float | Duration of speech in segment (excluding silence) | |
| | `word_timestamps` | list | Word-level timestamps (relative to speech-only audio) | |
| | `source_file` | string | Original audio filename | |
|
|
| ## VAD Processing |
|
|
| Audio segments are processed with Silero VAD to match faster-whisper inference: |
| - Silence is stripped from audio (only speech regions remain) |
| - Timestamps are relative to the concatenated speech audio |
| - This ensures training data matches inference behavior |
|
|
| ## Training Usage |
|
|
| For Whisper timestamp training, use the two-bucket approach: |
| - **Bucket A (50%)**: Use `text` - plain transcription without timestamps |
| - **Bucket B (50%)**: Use `text_ts` - transcription with Whisper timestamp tokens |
|
|
| ## Usage |
|
|
| ```python |
| from datasets import load_dataset |
| |
| dataset = load_dataset("Trelis/sample-dataset-test-neural-pack") |
| ``` |
|
|
| --- |
| *Prepared with [Trelis Studio](https://studio.trelis.com)* |
|
|
