aigc-x/Pronunciation-boldvoice

Pronunciation Assessment Dataset (BoldVoice + speechocean762)

Dataset for fine-tuning multimodal models on English pronunciation assessment.

Overview

Source	Samples	Audio Duration	Description
BoldVoice	38,182	10-20s	Non-native English learners, BoldVoice API annotations
speechocean762	5,000	1.6-20s	Public dataset, 5-expert scored, Mandarin speakers
Total	43,182

Schema

Column	Type	Description
audio	Audio (16kHz mono)	Speech recording
reference_text	string	Text the speaker intended to read
response	string	JSON annotation (see below)
source	string	boldvoice or speechocean762
duration	float	Audio duration in seconds
score	int	Overall pronunciation score (0-100)

Annotation Format (response JSON)

{
  "words": [
    {
      "word": "bear",
      "expected": ["B", "EH", "R"],
      "actual": ["B", "AH", "R"],
      "is_correct": false,
      "errors": [{"index": 1, "expected": "EH", "actual": "AH", "type": "substitution"}]
    }
  ],
  "summary": {
    "total_phonemes": 3,
    "correct_phonemes": 2,
    "error_count": 1,
    "score": 67
  }
}

• Phonemes in ARPAbet notation (no stress markers)
• Error types: substitution, deletion, insertion, mispronounced

Fine-tuning

pip install -r requirements.txt

# Fine-tune Gemma 4 E2B-it with LoRA
python finetune_gemma4_e2b.py --model google/gemma-4-E2B-it

# Custom settings
python finetune_gemma4_e2b.py --model /path/to/local/model --lr 1e-4 --epochs 2 --batch-size 2

Token Budget

Metric	Value
Median tokens/sample	1,197
p95 tokens/sample	2,674
Max tokens/sample	6,143