---
language:
  - he
license: mit
task_categories:
  - automatic-speech-recognition
tags:
  - speech-recognition
  - lyrics
  - hebrew
  - music
pretty_name: Caspi STT Benchmark
size_categories:
  - n<1K
---

# Caspi STT Benchmark

Hebrew speech-to-text (STT) evaluation dataset built from **Mati Caspi** songs: YouTube audio plus reference lyrics, packaged for Hugging Face.

## Dataset description

- **Audio**: 16 kHz mono WAV segments (one row per track or segment).
- **Text**: Reference transcript (lyrics or song title when lyrics are missing).
- **Metadata**: `id`, `youtube_id`, `title`, `song_name`.

Intended for STT benchmarking: compare model transcriptions to `text` (e.g. WER/CER).

## How to use

```python
from datasets import load_dataset, Audio

ds = load_dataset("ozlabs/caspi", split="train")
ds = ds.cast_column("audio", Audio(sampling_rate=16_000))

# Example row
ex = ds[0]
# ex["audio"] → decoded array; ex["text"] → reference transcript
```

## Source

- Audio: extracted from YouTube (playlists) at 16 kHz mono.
- Lyrics: provided manually or via Shazam; 

## License

MIT (or as specified in the repo). Audio and lyrics are used for research/evaluation; respect YouTube and lyric providers’ terms.