tuanamz/vi-sommelier-v0

vi-Sommelier v0

Vietnamese podcast dialogue dataset — sequential turn records with per-turn audio, transcripts, and quality signals. Produced by the podcast-pipeline (fork of NAVER Sommelier, adapted for Vietnamese).

Scale (v1)

• 502 source clips (~28.6 hours raw audio pulled from YouTube podcasts) — v1 adds 91 batch-1 clips
• 18,675 turn records across all clips
• ~24.8 hours clean 2-speaker audio after quality filters (estimated from v0 91.1% retention)
• 68 human-corrected gold turns (across 3 clips: vnlearn2_8-9xbU_learnvietnames, antruong_dxv_ep13_min, havesip_dh_148_hieuthuhai)
• 94.4 % clips are effectively dyadic (top-2 speakers dominate)

Data layout

Each clip has its own directory. Downstream consumers typically read <clip>.jsonl and load audio_path (relative to the JSONL) into their trainer.

<clip>/
├── <clip>.jsonl              # one JSON per turn — MAIN OUTPUT
├── <clip>/NNNNN_SPEAKER_XX.mp3   # per-turn audio (referenced by audio_path)
├── sortformer_probs.json + .npy  # per-frame speaker probability tensor
├── force_alignment.json      # word-level timing (wav2vec2-vi-250h)
├── attribution_check.json    # per-segment misattribution flag (ECAPA vs Sortformer)
├── bgm_check.json            # per-segment BGM detection (PANNs CNN14)
├── merged_speaker_map.json   # speaker-merge decisions
├── corrections.json          # OPTIONAL — human-corrected gold text + speaker remaps
└── turn_splits_candidates.json  # OPTIONAL — auto-splitter candidates for contested turns

Turn record schema

Selected key fields (see <clip>.jsonl for full):

{
  "conversation_id": "havesip_dh_148_hieuthuhai",
  "turn_id": "00027",
  "turn_index": 27,
  "speaker": "SPEAKER_00",
  "start": 68.48,
  "end": 69.12,
  "duration": 0.64,
  "audio_path": "havesip_dh_148_hieuthuhai/00027_SPEAKER_00.mp3",
  "audio_sample_rate": 16000,
  "asr": {
    "rover": "chưa",
    "rover_original": "Chưa. Em cảm giác là nó",
    "whisper": "chưa",
    "phowhisper": "em rất nhanh",
    "chunkformer": "em rất nhanh"
  },
  "gold": {
    "text": "chưa",
    "annotator": "manual_review",
    "annotated_at": "2026-07-13"
  },
  "flags": {
    "is_overlap": true,
    "is_short": true,
    "cospeech_frac": 0.778,
    "overlap_confidence": 0.9,
    "is_backchannel": false,
    "language_flipped": false
  },
  "contested_words": [
    {"word_index": 3, "word": "à", "start": 68.9, "end": 68.98,
     "reasons": ["mid_utterance_backchannel_token", "cospeech_verified"]}
  ],
  "words": [
    {"word": "chưa", "start": 68.48, "end": 68.62},
    ...
  ]
}

Signal fields (v0 novelty)

• flags.cospeech_frac — fraction of Sortformer 0.08 s frames in this turn with ≥ 2 speakers active. Direct signal of speaker overlap.
• flags.overlap_confidence — [0, 1] score combining cospeech_frac, short-duration bonus, and ASR disagreement. 537 turns have score ≥ 0.6.
• contested_words — VN backchannel tokens embedded mid-utterance where the speaker attribution is likely wrong (2,102 turns flagged, 13.2 %). Downstream training can mask loss at these positions.
• words — word-level FA timing (wav2vec2-base-vietnamese-250h) on 98.2 % of turns.

Pipeline

Details in podcast-pipeline repo (private):

1. pick_clips.py — yt-dlp fetches 4-min windows, 16 kHz mono WAV
2. main_original_ASR_MoE.py (--lang vi --vad --dia3 --ASRMoE --demucs) — Sortformer diarization, Silero VAD, PANNs BGM detection, Demucs vocals, 3-way ASR (Whisper-large-v3 + PhoWhisper-large + ChunkFormer-vi) with ROVER voting
3. check_bgm.py / verify_speaker_attribution.py / merge_speakers.py — per-clip sidecars
4. force_align_boundaries.py — wav2vec2 word-level timestamps
5. dump_sortformer_probs.py — save per-frame speaker probability tensor
6. extract_overlap_windows.py + gated MossFormer2_SS_16K — separate backchannels in Sortformer-detected co-speech windows (59 windows across 28 clips in v0)
7. normalize_to_hf_jsonl.py — assemble the JSONL export, apply all filters + gold corrections

Known limitations (v0)

• Half-duplex first: this v0 is optimised for half-duplex S2S SFT (Qwen2.5-Omni / LlamaOmni). Full-duplex (Moshi-style) v1 in progress — requires 2-channel per-speaker audio assembly using Sortformer probs + gated MossFormer2 splices.
• contested_words is a triage hint, not ground truth: 2,102 turns flagged, but only the 68 human-corrected turns have verified ground truth. Auto-splitter produces 327 candidate splits — needs professional annotation to promote to gold.
• Audio is 16 kHz mono (Whisper-native). Original YouTube source was 48 kHz Opus; future pulls preserve Opus via pick_clips.py -k for TTS/vocoder work.
• Speaker labels are per-clip (SPEAKER_00 / SPEAKER_01), not globally consistent. A speaker in clip A is unrelated to the same-labeled speaker in clip B.

License

Contact tuanamz for access + licensing terms.

Citation

@dataset{vi_sommelier_v0_2026,
  title = {vi-Sommelier v0: Vietnamese Podcast Dialogue Dataset},
  author = {Tuan Dinh},
  year = {2026},
  note = {v0 preview, 20.3 hours, half-duplex focus. Private dataset.}
}