hotdogs/thai-speech-20k

🎤 Thai Speech 20K — Thai Speech Dataset for VibeVoice Fine-tuning

ไทย/English — ชุดข้อมูลเสียงพูดภาษาไทย 20,000 ประโยค สำหรับ fine-tune โมเดล TTS
Source: Derived from Thanarit/Thai-Voice-Test7

🇹🇭 Dataset เสียงภาษาไทย 20,000 ตัวอย่าง ผู้พูด 1 คน (SPK_00001)
🇬🇧 20,000 Thai speech utterances, single speaker (SPK_00001)

---

🏷️ Source

Field	Detail
Original Dataset	Thanarit/Thai-Voice-Test7
Original Creator	Thanarit
Upstream Source	GigaSpeech2 (filtered Thai subset)
Speaker	SPK_00001 (single speaker)
Columns	ID, speaker_id, Language (th), audio, transcript, length, dataset_name, confidence_score
Preprocessing	Converted to JSONL (text: "Speaker 0: {transcript}", audio: WAV path)

📄 This dataset is a preprocessed version of Thanarit/Thai-Voice-Test7, which itself is a Thai subset of GigaSpeech2 filtered to a single speaker (SPK_00001). We convert the Parquet format to JSONL and add speaker prefix ("Speaker 0:") for VibeVoice fine-tuning.

---

📦 Contents

Files in this repo (16.5 MB)

File	Size	Description
train.parquet	1.1 MB	20k samples in Parquet format (snappy compressed)
train_speaker.parquet	1.1 MB	20k samples with Speaker 0: prefix, Parquet format
data/train.jsonl	3.5 MB	20,000 training samples (text + audio path)
data/train_speaker.jsonl	3.7 MB	20,000 samples with Speaker 0: prefix
data/audio_samples/	6.8 MB	50 sample WAV files (24 kHz mono)
README.md	~4 KB	This file

💡 Parquet format available via datasets library:

from datasets import load_dataset
ds = load_dataset("hotdogs/thai-speech-20k", data_files="train_speaker.parquet")

Full audio (not uploaded)

Item	Detail
Total audio	2.1 GB (20,000 WAV files)
Format	WAV, 24 kHz, 16-bit mono
Duration	~11 hours total
Per-sample	65–620 KB, ~1–3 seconds each

---

📋 Data Format

JSONL with fields:
  text  — "Speaker 0: <Thai text transcript>" (UTF-8)
  audio — Path to WAV file (24 kHz, mono)

Example:

{"text": "Speaker 0: สวัสดีครับ วันนี้อากาศดีมาก", "audio": "audio/sample_000000.wav"}

Upstream format (Thanarit/Thai-Voice-Test7)

Parquet with columns:
  ID, speaker_id (SPK_00001), Language (th), audio, transcript, 
  length, dataset_name (GigaSpeech2), confidence_score

---

🎯 Intended Use

This dataset was prepared specifically for fine-tuning the microsoft/VibeVoice-1.5B model for Thai text-to-speech with speaker conditioning. It is intended for:

• Text-to-Speech (TTS): Fine-tuning TTS models for Thai language
• Speaker Adaptation: Single-speaker voice cloning/personalization
• Low-resource TTS: Thai TTS research with limited data

---

⚠️ Limitations

• Single Speaker: Only one speaker (SPK_00001) — may not generalize to multi-speaker scenarios
• Private Audio: Audio files are not publicly available for privacy reasons
• Domain: General conversational Thai only — no domain-specific vocabulary
• No Evaluation Split: All 20k samples used for training; evaluation done separately

---

📎 Related

• Upstream Dataset: Thanarit/Thai-Voice-Test7 — original Thai subset from GigaSpeech2
• Trained Model: hotdogs/vibevoice-1.5b-thai-tts-lora — LoRA adapter trained on this dataset

---

🙏 Credits

• Original Dataset: Thanarit — Thai-Voice-Test7 (from GigaSpeech2)
• Preprocessing: UKA — JSONL conversion + speaker prefix for VibeVoice
• Purpose: Fine-tuning VibeVoice for Thai TTS
• Timestamp: 2026