OD3: Open Directed Dialogue Dataset

OD3 (Open Directed Dialogue Dataset) is designed to allow the community to explore further research into leveraging flawed conversational interactions to improve model performance. OD3 is a collection of 63K conversations (600K turns, 1,172 hours of audio) drawn from existing natural language task-oriented dialog datasets, and augmented with synthetic audio. OD3 is further augmented with turns containing repeats and rephrases of previous failed utterances.

For more details, check out the paper.

• Paper: Task Oriented Dialogue as a Catalysis for Self-Supervised Automatic Speech Recognition (ICASSP 2024)
• Original repository: amazon-science/amazon-od3
• Contact: davidchan@berkeley.edu
• License: CC-BY-NC-SA-4.0

Dataset statistics

Computed directly from the released annotation files:

Split	Conversations	Turns	Audio clips	Repeat/rephrase turns	Distinct voices
train	47,702	472,347	638,851	9,199	82,527
validation	6,437	64,631	95,778	1,375	12,470
test	8,821	85,610	126,945	1,816	17,108
Total	62,960	622,588	861,574	12,390	—

Seed dataset composition (conversations)

Source dataset	train	validation	test	total
DSTC11-Track-5	25,579	4,085	5,342	35,006
MultiWOZ 2.2	8,425	1,000	999	10,424
SIMMC 2.1	7,306	563	1,687	9,556
NOESIS	3,967	487	489	4,943
KVRET	2,425	302	304	3,031

Data structure

Each line of a .jsonl file is one conversation:

{
  "sample_id": "b4bbb5b2-c075-4a8b-a953-8b14cd861ead",
  "source_dataset": "DSCT11-Track-5",
  "is_llm_generated": false,
  "meta": {"knowledge": null, "target": false},
  "turns": [
    {
      "text": "I am looking for a certain hotel...",
      "speaker_id": "0d6bb10a-76da-4f7e-bd90-2416a4449a8c",
      "turn_id": "aa0a0989-b545-4620-a7d9-47e7e0abf82d",
      "is_agent": false,
      "turn_is_repeat_rephrase": false,
      "meta": {
        "is_induced_by_repeat_rephrase": false,
        "is_machine_generated": false,
        "repeat_rephrase_type": null
      },
      "audio": {
        "25218975": {
          "path": "dev/b4/b4bbb5b2-.../aa0a0989-...+25218975.wav",
          "voice_id": "25218975",
          "universal_agent": false,
          "asr_transcript": " I am looking for a certain hotel...",
          "normalized_ground_truth": "i am looking for a certain hotel...",
          "normalized_asr_transcript": "i am looking for a certain hotel..."
        }
      }
    }
  ]
}

Fields

Conversation level

• sample_id — unique conversation identifier (UUID).
• source_dataset — the seed dataset this conversation was drawn from.
• is_llm_generated — whether the conversation itself was LLM-generated.
• meta — source-specific metadata (e.g. knowledge, target).

Turn level

• text — the ground-truth utterance text.
• speaker_id — synthetic speaker identity, stable within a conversation.
• turn_id — unique turn identifier (UUID).
• is_agent — true for agent turns, false for user turns.
• turn_is_repeat_rephrase — whether this turn is a repeat/rephrase of a previous failed turn.
• meta.repeat_rephrase_type — the kind of repeat/rephrase, when applicable.
• audio — map of voice_id → audio rendition.

Audio rendition

• path — path within the audio tree: <split>/<first-2-chars-of-sample_id>/<sample_id>/<turn_id>+<voice_id>.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

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.

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
<split>/<sample_id>.json	the complete conversation record — every turn, all metadata (including DSTC11 knowledge grounding), plus split, n_audio, and audio_files
<split>/<sample_id>.000.wav	clips, numbered in turn order then voice order
<split>/<sample_id>.001.wav	…

The WebDataset __key__ is <split>/<sample_id>.

audio_files maps each wav field to its place in the dialogue, which is how you attach a clip to its turn:

"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

@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. This repository mirrors the dataset originally released at 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.