---
language:
- ar
license: cc-by-4.0
task_categories:
- text-to-speech
tags:
- arabic
- tts
- single-speaker
- snac
- orpheus
- high-quality
pretty_name: Arabic Male Host — Single Speaker TTS (15h, STOI≥0.95)
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
dataset_info:
  features:
  - name: clip_id
    dtype: string
  - name: text
    dtype: string
  - name: audio_tokens
    list: int64
  - name: token_count
    dtype: int64
  - name: duration_s
    dtype: float64
  - name: duration_snac_s
    dtype: float64
  - name: squim_stoi
    dtype: float64
  - name: squim_pesq
    dtype: float64
  - name: squim_si_sdr
    dtype: float64
  - name: speaker_id
    dtype: string
  - name: speaker_gender
    dtype: string
  - name: channel
    dtype: string
  - name: lang
    dtype: string
  - name: audio
    dtype:
      audio:
        sampling_rate: 24000
  - name: sim_to_ref
    dtype: float64
  splits:
  - name: train
    num_bytes: 3675630204
    num_examples: 5459
  download_size: 3674884410
  dataset_size: 3675630204
---

# Arabic Male Host — Single Speaker TTS Dataset

**High-quality, single-speaker Arabic TTS dataset.** 3,769 clips / 15.3 hours of one consistent Arabic male voice, extracted from thmanyahpodcasts (Thamanyah) — one of the largest Arabic podcast networks.

This is a **clean, ready-to-train** dataset for Arabic TTS systems (Orpheus, VITS, etc.). All clips come from the same speaker, verified via speaker embedding cosine similarity (0.978 across source batches).

## Speaker

- **Gender**: Male
- **Language**: Arabic — Gulf/Saudi dialect, natural conversational style
- **Source**: thmanyahpodcasts YouTube channel (videos 169–368)
- **Identity verified**: Speaker cosine similarity 0.978 across both batches — same person throughout

## Dataset Stats

| Stat | Value |
|------|-------|
| Clips | 3,769 |
| Total duration | 15.31 hours |
| Sample rate | 24 kHz mono |
| Clip length | 3–25 seconds |
| Min STOI quality | ≥ 0.95 |
| SNAC tokens | ✅ Included |
| WAV audio | ✅ Included |

## Columns

| Column | Description |
|--------|-------------|
| `clip_id` | Unique clip identifier |
| `text` | Arabic transcription (Gemini 2.0 Flash, includes [laughter] [pause] tags) |
| `audio` | WAV audio at 24 kHz mono |
| `audio_tokens` | SNAC-encoded token list (snac_24khz, 7-token interleaved stripe) |
| `token_count` | Number of SNAC tokens |
| `duration_s` | Clip duration in seconds |
| `squim_stoi` | Intelligibility score — all clips ≥ 0.95 |
| `squim_pesq` | PESQ perceptual quality score |
| `squim_si_sdr` | SI-SDR score |
| `speaker_gender` | `male` |
| `channel` | `thmanyahpodcasts` |
| `lang` | `ar` |

## Quality Pipeline

1. **VAD segmentation** — Silero-VAD, 3–25s clips
2. **Quality gate** — torchaudio SQUIM STOI ≥ 0.95 (filters music, noise, low-quality audio)
3. **Speaker clustering** — WeSpeaker embeddings, agglomerative clustering (threshold 0.80)
4. **Transcription** — Gemini 2.0 Flash via OpenRouter, 50 parallel workers
5. **Language filter** — Arabic-only clips retained

## SNAC Token Format (for Orpheus training)

- Vocoder: `hubertsiuzdak/snac_24khz`
- 7-token interleaved stripe per frame: `[L0, L1, L1, L2, L2, L2, L2]`
- Mapped token ID formula: `128266 + (slot % 7) * 4096 + code`
- Special tokens: SOA=128257, EOA=128258, EOT=128009
- Training format: `<custom_token_1>voice_name: Text<|eot_id|> [SNAC TOKENS] <custom_token_2>`

## Related Datasets

- Full Arabic multi-speaker SNAC datasets (74k clips, 197h total):
  - `vysakh25/youtube-arabic-src0-68-snac`
  - `vysakh25/yt-arabic-src69-168-snac`
  - `vysakh25/yt-arabic-src169-268-snac`
  - `vysakh25/yt-arabic-src269-368-snac`