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.

Why this exists

This project's other 6 training-mix sources each have real audio, but none of them contain genuine multi-turn listen-then-speak structure -- every existing <listen>/<speak> example is a single isolated exchange. This was a known, explicitly flagged gap (the model has never seen a sustained back-and-forth conversation where it has to track what it previously "heard" and "said" across several turns). DailyTalk is small (~20.5h vs. e.g. the voiceclap-data emolia subset's ~9,750h) but is the one dataset found with the right STRUCTURE, under a genuinely permissive license -- most comparable multi-turn spoken-dialogue datasets found during sourcing (SALT-Research/DeepDialogue-orpheus, ICTNLP/InstructS2S-200K) are CC-BY-NC-4.0 and excluded per this project's permissive-license-only rule.

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

• M (primary, new capability): the WHOLE conversation as one chatml example, turns sorted by turn_id and alternated by position parity -- even position -> <|im_start|>user + <listen> (downsampled), odd position -> <|im_start|>assistant + <speak> (full). All turns of one conversation chained into a single example (avg ~9.4 turns, max 20 observed -- comfortably under seq_length=8192, no chunking needed).
• A (single-turn TTS): [Speaker: 0 or 1] Say: "{text}" -> <speak>.
• B (single-turn ASR): <listen>{audio}</listen> + a question-pool template -> {text} (the real transcript).

All 3 tasks reuse ONE SNAC encode_speak() pass per turn (7 tok/12.5Hz frame, same offsets/scheme as this project's other SNAC sources laion_emotional_roleplay/voiceclap_emolia), downsampled to <listen> (3 tok/frame, keep L0/L1a/L1b) for whichever turns land in the USER role -- no separate encode pass.

Format (Task M shown, chatml tier -- every row in this release is chatml,

there is no base/USER:-ASSISTANT: tier for this source)

<|im_start|>user
<listen> ...tokens... </listen><|im_end|>
<|im_start|>assistant
<think>
</think>
<speak> ...tokens... </speak><|im_end|>
<|im_start|>user
<listen> ...tokens... </listen><|im_end|>
<|im_start|>assistant
<think>
</think>
<speak> ...tokens... </speak><|im_end|>
... (repeated for every turn in the conversation)

Base

keonlee9420/DailyTalk (CC-BY-SA 4.0), re-recorded from DailyDialog (English, open-domain). Downloaded via the eustlb/dailytalk HF mirror (no license tag of its own -- treated as inheriting the original's CC-BY-SA-4.0). Audio SNAC-tokenized with tokenizer-vla-qwen3-v2.