You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Duplex training metadata

Join by globally unique id. Columns: id, source, language, scenario, frame_ms, num_frames, start_switches, yield_switches, label_source, review_status, details. details is a JSON string with source metadata and audio checksums. original-supervision.zip retains the full original 10-head targets/masks, original versus corrected labels, acoustic/VAD arrays, synthesis metadata and preparation code.

Matching repository: https://huggingface.co/datasets/giangndm/duplex-training-audio

Public repository with manual access approval. Request access to both repositories.

Verified: contents and alignment

  • synthetic: 768 scripted paired scenes, train/dev/test 384/192/192; 20.48 s each.
  • utterances: 160 original ZeroTTS clips, train/dev/test 80/40/40; 16 English clips.
  • A/B share the original timeline. No independent silence trimming is applied.
  • Every Opus file was decoded and its sample count checked against its source.
  • Audio is lossy Opus, 24 kbps target VBR, mono, 20 ms packets. Decode/resample both tracks to 16 kHz for training; an Opus decoder may expose 48 kHz by default.
  • Synthetic duration is 4.369 composed hours from 275.68 seconds of unique clips. Reused clips and repeated wording are not independent conversation hours.

Proposal: label contract

start_switches and yield_switches are independent timelines. Each item is {frame: int, value: 0 | 1 | null} and applies over [frame, next_frame), ending at num_frames. 0 means negative, 1 positive, null means mask out of loss. Empty timelines mean no action labels (isolated utterances). Frames are 80 ms. Do not fill unknown regions with zero. START means begin the main answer; YIELD means stop playback/yield. Simultaneous head predictions need runtime arbitration.

Synthetic labels are scripted intent, cold-start-v1. Open Yap labels are observational timing proxies derived from Silero VAD, NOT verified stop intent. All rows have review_status=pending: no human annotation is claimed. In semantic training filter label_source=script; use timing-proxy only as an explicit weak supervision baseline. No BACKCHANNEL action or human USER_TURN labels are invented. Transcripts, seeds and source metadata are in details or the original archive.

Verified: source and limitations

Open Yap audio is not duplicated here; use the upstream dataset separately.

ZeroTTS model (MIT model card), revision pinned: https://huggingface.co/zeroweight-ai/ZeroTTS/tree/c2bfbd67dc648cac455077333f7cf5c18a2e3bb4 Synthetic voices are built-in voice packs. Pronunciation is not human-validated; English Stop/Okay is not a validated standalone English TTS benchmark. The model license does not establish an independently verified license for every voice/output. No additional blanket license is asserted for the mixed-source collection. Source cards inspected 2026-11-13.

Speaker/voice identities are disjoint across splits; synthetic wording and scene templates are shared. Opus changes the waveform: old encoder feature caches are not interchangeable with features extracted from the distributed Opus files.

Load and join

from datasets import load_dataset, Audio, concatenate_datasets

config, split = "synthetic", "train"  # also: utterances
audio = load_dataset("giangndm/duplex-training-audio", config, split=split, token=True)
meta = load_dataset("giangndm/duplex-training-meta", config, split=split, token=True)
assert audio["id"] == meta["id"]  # Always verify; never join only by row number.
joined = concatenate_datasets([audio, meta.remove_columns("id")], axis=1)
joined = joined.cast_column("user_a", Audio(sampling_rate=16000))
joined = joined.cast_column("user_b", Audio(sampling_rate=16000))

Use a token approved for BOTH repositories. Pin both commit revisions for training. The utterances config has null user_b and no action supervision. Filter metadata language == "en" for the 16 English Stop/Okay utterances; no full English synthetic two-channel dialogue corpus is claimed.

daily-talk-v3

Synthetic Vietnamese daily conversations generated by ZeroTTS-family PyTorch and checked by Gipformer. Public repository with manual access approval. Labels are scripted and pending human review. Natural clause boundaries model backchannel continue, pause/resume and genuine interruption stop. Audio stores exactly id/user_a/user_b with embedded Opus; metadata joins by id. Streaming paired uploads have receipts in manifests; pin both recorded revisions before training. Synthetic variants include authored scenario grammars; counts of unique accepted speech and scene timeline are reported separately. No hours are inferred from text, silence or repeated bank clips.

daily-talk-authored-v1

Synthetic speech; metadata remains pending human review. Audio is embedded Opus in id/user_a/user_b columns; join metadata by id. Pin matching manifest revisions before training.

Downloads last month
8,667