--- license: cc-by-nc-sa-4.0 language: - en task_categories: - automatic-speech-recognition - text-generation tags: - dialogue - task-oriented-dialogue - speech - asr - conversational - repeats-and-rephrases pretty_name: 'OD3: Open Directed Dialogue Dataset' size_categories: - 10K///+.wav`. - `voice_id` — the synthetic voice used (cloned from CommonVoice). - `asr_transcript` — transcript produced by running ASR over the synthesized audio. - `normalized_ground_truth` / `normalized_asr_transcript` — normalized text for WER computation. - `word_error_rate` — WER of this rendition's ASR transcript against the ground truth. - `nbest` — up to 8 n-best ASR hypotheses for the clip. The gap between `normalized_ground_truth` and `normalized_asr_transcript` is what makes OD3 useful for studying self-supervised ASR improvement from dialogue signals. ## Usage ### Annotations only ```python from datasets import load_dataset ds = load_dataset("davidchan/od3") # train / validation / test print(ds["train"][0]["sample_id"]) ``` ### Conversations with audio The audio is published as **conversation-grained WebDataset shards** under `audio/`: **one WebDataset sample is one complete conversation**, not one clip. ```python import json, webdataset as wds url = "https://huggingface.co/datasets/davidchan/od3/resolve/main/audio/train-00000.tar" for sample in wds.WebDataset(url, shardshuffle=False): conv = json.loads(sample["json"]) # the whole conversation print(conv["sample_id"], len(conv["turns"]), "turns,", conv["n_audio"], "clips") for field, meta in conv["audio_files"].items(): wav = sample[field] # raw RIFF/WAVE bytes turn = conv["turns"][meta["turn_index"]] print(" ", turn["text"][:50], "| voice", meta["voice_id"], "|", len(wav), "bytes") break ``` ## Audio ``` audio/train-00000.tar ... train-00099.tar 100 shards audio/dev-00000.tar ... dev-00014.tar 15 shards audio/test-00000.tar ... test-00019.tar 20 shards ``` Each shard holds whole conversations (~1GB per shard). A conversation is **never split across shards**. Within a sample: | File | Contents | | ---- | -------- | | `/.json` | the complete conversation record — every turn, all metadata (including DSTC11 `knowledge` grounding), plus `split`, `n_audio`, and `audio_files` | | `/.000.wav` | clips, numbered in turn order then voice order | | `/.001.wav` | … | The WebDataset `__key__` is `/`. `audio_files` maps each wav field to its place in the dialogue, which is how you attach a clip to its turn: ```json "audio_files": { "000.wav": {"path": "train/e9/.../3da7d281-...+24797540.wav", "turn_index": 0, "turn_id": "3da7d281-...", "voice_id": "24797540"} } ``` ### Things to know - **`sample_id` is unique only within a split.** 486 conversations appear in *both* `dev` and `test` — same turns and text, resynthesized with different voices. Key on `(split, sample_id)`, and be aware of this overlap if you evaluate on both splits. - **A turn may have 0, 1, or 2 renditions** (different synthetic voices): 299,136 turns have one, 281,219 have two, and 42,233 are text-only. Conversations run up to 72 clips. - **Text-only conversations are not in the shards.** 4,649 `train` conversations have no audio at all; they appear in `od3_train.jsonl` only. The shards cover 58,311 conversations and all 861,574 clips. - **`dev_mini` is not included.** The original `audio.tar.gz` contains a small `dev_mini/` directory (5 conversations) holding an extra voice rendering of a handful of `dev` conversations. No annotation file references those paths, and the renditions have no `voice_id`, ASR transcript, or WER, so they are omitted here. The shards contain exactly the 861,574 clips described by the annotations. - **Use the `webdataset` reader for audio.** Because samples have a variable number of `.wav` fields, the `datasets` WebDataset builder (which infers one fixed schema) is a poor fit here; read the shards with `webdataset` and use `load_dataset` for the JSONL annotations. ### Original tarball The audio was originally distributed as a single `audio.tar.gz` (93.3 GB) from CloudFront. That object exceeds CloudFront's 50 GiB per-object limit, so a plain `curl -O` returns **HTTP 400**; only bounded HTTP range requests succeed. The shards here contain byte-for-byte the same WAV files, regrouped by conversation. ## Citation ```bibtex @inproceedings{chan2023domain, title={Task Oriented Dialogue as a Catalysis for Self-Supervised Automatic Speech Recognition}, author={Chan, David M and Ghosh, Shalini and Tulsiani, Hitesh and Rastrow, Ariya and Hoffmeister, Bj{\"o}rn}, booktitle={ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, year={2024}, organization={IEEE} } ``` ## License and attribution Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved. OD3 is released under [CC-BY-NC-SA-4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/legalcode). This repository mirrors the dataset originally released at [amazon-science/amazon-od3](https://github.com/amazon-science/amazon-od3). OD3 is derived from DSTC11 (Track 5), KVRET, MultiWOZ 2.2, NOESIS, and SIMMC 2, and uses the Mozilla CommonVoice corpus (`cv-corpus-14.0-2023-06-23`) for voice cloning. Each seed dataset remains subject to its own license.