Meddies/meddies-asr-synth-dialog

Meddies ASR — Synthetic Dialog Speech (vi/en/zh)

Synthetic doctor–patient consultations for ASR training, generated with Fish Audio TTS (s2.1-pro-free, 16 kHz mono FLAC) over curated conversational voice registries (en 75 / zh 165 voices) with 22 compositional scenario profiles driving expressiveness (disfluencies, emotion palettes, speed jitter, age-band casting, pause pacing). Target: 1,000 h per language.

Configs

• vi_dialog: Vietnamese clinical dialogue turns with embedded 48 kHz FLAC. The Hugging Face viewer can preview and play individual samples directly. Group by conv_id, then sort by turn and unit_index to reconstruct a full dialogue; parent_row_id groups units from the same original turn.
• {lang}_dialogs (primary, 2026-07-03 →): one row per full consultation — long-form audio with per-turn alignment.
• {lang}_utterances (frozen, ~102 h): one row per turn, from the pre-dialog-format pipeline. Still valid training data; no longer growing.

Dialog row schema

column	meaning
audio	full-consult 16 kHz mono FLAC (embedded), pause-sampled turn gaps
text	clean transcript, \n-joined, role-prefixed (doctor: … / patient: …)
segments	list of {start_s, end_s, role, text, voice_id} per turn — last end_s == duration_s exactly; slice these for 15–30 s training windows
dialog_id, lang, n_turns, duration_s	identity + shape
profile	scenario profile (22 named combos; baseline = flat pre-profile generation) — the WER-ablation lever

Transcripts are always clean text_label — emotion/paralanguage markup is spoken but never transcribed (deliberate disfluency-robustness for ASR).

state.json at the repo root is the generation cursor (per-lang offsets + shipped hours) — infrastructure, not data.

Generation: prosody speed 0.9 ± profile jitter (clamped 0.7–1.15), temperature 0.55, top_p 0.6, 8 req/s pooled. Internal / private. Built by the meddies-asr synthgen pipeline.