dailytalk-flattened

SNAC-tokenized, flattened + augmented build of DailyTalk (original: keonlee9420/DailyTalk, CC-BY-SA 4.0 -- downloaded via the eustlb/dailytalk HF mirror). 2,541 real 2-speaker dialogues re-recorded from the open-domain text dataset DailyDialog, purpose-built for conversational TTS conditioned on dialogue history -- ~20.5 hours total audio, 23,773 turns.

v3 update (2026-08-11): full-listen + separate vocab + format diversity

Replaces the earlier build with the converged design validated by an ablation study (2/3/5) on <listen>/<speak> token format:

• Full <listen> content: 7 tok/12.5Hz frame (not truncated to 3).
• Separate vocab for <listen>: every listen-side SNAC id is shifted +1,000,000 vs the identical <speak> code, so the two roles never share an embedding -- this recovered most TTS reliability that sharing ids had cost, per the ablation comparison, while keeping most of the ASR/comprehension gain over the pre-ablation baseline.
• 4-way format diversity on single-turn A/B records (Instruction/ Response, User/Assistant, USER/ASSISTANT, Qwen chatml -- hash-picked per record id, see data_prep/audio_ablations/ablation5_format_diversity/style_wrap.py). The multi-turn M task stays chatml (format-diversity is single-turn-only by design).

Requires tokenizer-vla-qwen3-v3 (or newer) -- the shifted <snac_{N+1000000}> ids used by <listen> are not atomic in earlier tokenizers.

Why this exists

This project's other training-mix sources each have real audio, but none of them contain genuine multi-turn listen-then-speak structure -- every other existing <listen>/<speak> example is a single isolated exchange. DailyTalk is small (~20.5h) but is the dataset found with the right STRUCTURE under a genuinely permissive license.

Tasks (3 per conversation/turn, non-LLM -- pure field reuse + templates)

• M (primary): the WHOLE conversation as one chatml example, turns alternated by position parity -- even position -> <|im_start|>user + <listen> (full, shifted vocab), odd position -> <|im_start|>assistant
  • <speak> (full, unshifted). All turns of one conversation chained into a single example (avg ~9.4 turns, max 20 observed).
• A (single-turn dialogue continuation, format-diverse): USER = the real plain text of the PRECEDING turn, ASSISTANT = <speak> of the CURRENT turn's real reply. Only emitted for turn_idx > 0 (the first turn of a conversation has no prior turn to prompt with), so there are 2,541 fewer A records than B records (21,232 vs 23,773). Fixed 2026-08-11 (Huu, Discord 2026-08-07): the original version used [Speaker: N] Say: "{text}" -> <speak>{same text} framing, which no real dialogue would ever phrase, and collided with this project's genuine TTS-instruction tasks (emolia/voiceclap's Task A/F) where Say: "X" really does mean "speak this exact text on command" -- here it's a real conversational reply, not an echo, so that phrasing was actively misleading for training.
• B (single-turn ASR, format-diverse): <listen>{audio}</listen> + question-pool -> {text} (real transcript). Unaffected by the A fix.

Base

keonlee9420/DailyTalk (CC-BY-SA 4.0), re-recorded from DailyDialog (English, open-domain). Audio SNAC-tokenized (hubertsiuzdak/snac_24khz, full 3-level 7 tok/frame), <listen> ids shifted +1,000,000 vs <speak>'s.