tungnguyenlam/vietnamese-acoustic-boundary-verifier-data

Vietnamese Acoustic Boundary & Speaker Purity Dataset (Gemini 3.8 Flash Distilled)

This dataset contains 202 curated Vietnamese audio samples with fine-grained acoustic boundary annotations distilled from Google Gemini 3.8 Flash (thinkingLevel="LOW").

It is specifically designed to train and evaluate multimodal models (e.g., Gemma 4 E4B Audio) on acoustic quality control for speech synthesis and speaker diarization pipelines.

Dataset Structure

Each sample is formatted with the prompt used during distillation:

{
  "audio": "audio/synth_clip_coda_subtle_01.wav",
  "prompt": "...[Acoustic Overlap & Boundary Verification Prompt]...",
  "target_json": {
    "speaker_purity": "pure",
    "word_completeness": "clipped",
    "boundary_issue": "clipped_word_end",
    "failure_codes": ["clipped_word_end"],
    "reason": "Âm tiết cuối bị cắt cụt đột ngột làm mất đuôi thanh điệu..."
  },
  "decision": "reject",
  "source": "coda_subtle"
}

Acoustic Categories Included

1. 1.Clean Passes (`clean`): Natural conversational turns with intact acoustic onsets and codas.
2. 2.Subtle Coda Cuts (`coda_subtle` ~80ms): Cuts the trailing tonal contour or offglide without deleting the entire vowel.
3. 3.Moderate Coda Cuts (`coda_moderate` ~180ms): Truncates final consonants/vowels ("lẹm chữ").
4. 4.Subtle Onset Cuts (`onset_subtle` ~90ms): Shaves initial consonant burst/plosives.
5. 5.Severe Onset Cuts (`onset_severe` ~220ms): Cuts the initial syllable onset completely.
6. 6.Cross-Talk / Bleed (`tail_bleed`, `head_bleed` 250–350ms): Leading or trailing vocal intrusions from secondary speakers.

Teacher Model

• —Model: Google Gemini 3.8 Flash
• —Reasoning Effort: LOW