---
tags:
- audio
- speech
- whisper
- dataset
---
# test-draft-asr-e2e

Speech dataset prepared with Trelis Studio.

## Statistics

| Metric | Value |
|--------|-------|
| Source files | 1 |
| Train samples | 4 |
| Total duration | 2.1 minutes |

## Columns

| Column | Type | Description |
|--------|------|-------------|
| `audio` | Audio | Audio segment (16kHz) - speech only, extracted from aligned regions |
| `text` | string | Plain transcription (no timestamps) |
| `text_ts` | string | Transcription with Whisper timestamp tokens |
| `preconditioning` | string | Previous segment's plain text (empty for first segment of each source file) |
| `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` | string | Word-level timestamps as JSON (relative to speech-only audio) |
| `source_file` | string | Original audio filename |
| `language` | string | ISO 639-1 language code for Whisper language token (e.g., `en`, `el`) |

## Speech Segmentation

For `.txt` transcripts, CTC alignment produces word-level timestamps that define speech
boundaries. The full audio is treated as a single speech region and split at word boundaries
into ~20s chunks. For `.srt`/`.vtt` files, transcript timestamps define segments.
Timestamps are relative to the extracted speech audio within each chunk.

## Training Usage

### 2-bucket approach (default)
- **Bucket A (50%)**: Use `text` column - plain transcription without timestamps
- **Bucket B (50%)**: Use `text_ts` column - transcription with Whisper timestamp tokens

### 4-bucket approach (with preconditioning)
- **Bucket A (25%)**: `text` only (no timestamps, no preconditioning)
- **Bucket A' (25%)**: `preconditioning` + `text` (previous segment context, no timestamps)
- **Bucket B (25%)**: `text_ts` only (timestamps, no preconditioning)
- **Bucket B' (25%)**: `preconditioning` + `text_ts` (previous segment context + timestamps)

Preconditioning prepends the previous segment's text to teach the model conversational continuity.
Bucket ratios are configurable at training time.

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("Trelis/test-draft-asr-e2e")
```

---
*Prepared with [Trelis Studio](https://studio.trelis.com)*