JonesLin/nano-duplex-data-v1

nano-duplex data v1

English speech alignment and a first natural-conversation collection for nano-duplex research. All audio is embedded in Parquet as PCM16 FLAC at 24 kHz. No external audio paths are required.

Configuration	Split	Examples	Audio hours	Speakers
alignment	train	26,478	32.377	247
alignment	validation	4,708	5.817	40
alignment	test	3,624	4.428	38
interaction	train	946	15.240	57
interaction	validation	122	1.978	11
interaction	test	116	1.877	11

Sources and processing

Alignment: LibriTTS-R, by Yuma Koizumi and colleagues. The complete official train_clean_100, dev_clean, and test_clean archives passed the official MD5 checks before extraction. Keep complete mono 24 kHz PCM16 utterances from 1 through 10 seconds with normalized transcripts. Preserve official splits and original IDs. Reject missing transcripts, out-of-range durations, duplicate audio, and entirely silent clips. WAV-to-FLAC conversion is lossless; original WAV and decoded PCM SHA256 are recorded. Original and normalized text are included; the source transcript hash refers to the original normalized-text file, before trimming its surrounding whitespace. RMS and clipping metrics are diagnostics, not human quality judgments. This is a filtered subset, not all hours implied by the upstream archive names. Citation: Koizumi et al., "LibriTTS-R: A Restored Multi-Speaker Text-to-Speech Corpus," 2023.

Interaction: otoearth/otoSpeech-full-duplex-processed-141h. Use eight complete upstream tar shards, selected by lexicographic filename order, at revision c635c29ad083570fe2cadbafc75770e76eded2da. Every shard passed its pinned Hub SHA256. The selection contains 73 complete sessions. Resample each complete stereo session from 44.1 kHz to 24 kHz using SciPy polyphase filtering, then split into consecutive 60-second windows (including the shorter final window). Both channels retain the shared clock. Preserve overlapping speech and silence, and re-mute the original redacted intervals after resampling. The two speaker_ids are ordered by audio channel. Source-relative timestamps allow exact reassembly. Speaker-connected components stay within one split; there is no reported speaker-ID or session overlap across train/validation/test. Speaker identities beyond the IDs reported by the source have not been independently verified. Omit optional surveys and demographic profiles. Citation: otoearth, "otoSpeech-full-duplex-processed-141h," 2026.

Scope

Natural conversations have no assistant tool traces, cancelled playback plans, certified prefix-causal action targets, or transcripts in this release. causal_verified=false and has_transcript=false are explicit. These recordings support interaction preparation and analysis; they are not ready-made Stage B-D causal trajectories. No synthetic speech is substituted for real recordings. Tool, vision, and motion supervision are outside this release.

codec-pilot/ contains real Mimi ASR features and eight-codebook TTS targets for 773 training utterances (57.301 minutes), drawn from alignment/train. This pilot has no independent evaluation split. Its JSONL manifests reference tensor files packed in tensors.tar; cache-index.json records every member's size and SHA256. The server entry expands and verifies them automatically. The full alignment and interaction collections are processed audio; only this pilot is codec-encoded. The exact Mimi repository, revision, checkpoint digest and cache dimensions are recorded. Tensor files must be loaded with torch.load(..., weights_only=True) as nano-duplex does. The provided codec-pilot/asr.toml and codec-pilot/tts.toml select matching pinned assets.

Loading

from datasets import load_dataset, Audio

alignment = load_dataset("JonesLin/nano-duplex-data-v1", "alignment", token=True)
interaction = load_dataset("JonesLin/nano-duplex-data-v1", "interaction", token=True)
# Decode with soundfile without requiring torchcodec:
audio = alignment["train"].cast_column("audio", Audio(decode=False))[0]["audio"]
from io import BytesIO
import soundfile as sf
pcm, sample_rate = sf.read(BytesIO(audio["bytes"]))

From the nano-duplex project checkout, install requirements-speech.txt and requirements-data.txt, then use the server entry (a copy is included at tools/hf_training.py; it belongs in the project's nano_duplex package):

# Download only the small pre-encoded pilot, verify and expand it, then train ASR/TTS.
python -m nano_duplex.hf_training --profile pilot --steps 400

# Full Stage A: download alignment data, encode each split once, train, evaluate.
python -m nano_duplex.hf_training --profile full --nproc 4 --steps 20000

The first command resolves a concrete Hub commit and writes dataset-lock.json. Alternatively pass --revision with an exact commit. --snapshot-dir, --work-dir, and --run-dir select server storage. --phase prepare stops before training; --phase train --resume reuses data caches and resumes existing task checkpoints. An unstarted second task starts fresh. Use --dry-run to inspect commands without network access or file writes. H200 distributed execution has not been run here; 20,000 steps is a starting budget, not a measured convergence requirement.

The full profile still performs one Mimi encoding pass over the complete alignment collection on first use. Subsequent runs reuse checked prepared manifests. Use tools/materialize_alignment.py separately if only mono WAV/raw JSONL export is needed. The interaction configuration is available through load_dataset, but is not fed into the causal duplex trainer without the additional labels described above. Changing a dataset manifest requires a new training run, not resuming an old optimizer state.

Provenance and license

Both source corpora are provided under CC BY 4.0. This derivative retains that license and the source attributions above. The source-provided gated dataset access was obtained through the collecting user's own Hugging Face account. Source receipts, split statistics, exclusions, processing parameters, and full validation results are in reports/. MANIFEST.json inventories the exact release files with SHA256. The validation reads every packaged audio example, checks its PCM hash, duration, channel count, session continuity, redaction masks, and split isolation. No speech-quality or conversational-performance claim follows from those structural checks.