README.md

metadata

language:
  - en
license: cc-by-4.0
task_categories:
  - automatic-speech-recognition
tags:
  - speech
  - transcription
  - conversation
  - turn-taking
  - seamless
  - canary
size_categories:
  - 1M<n<10M
dataset_info:
  features:
    - name: audio_filename
      dtype: large_string
    - name: offset
      dtype: float64
    - name: duration
      dtype: float64
    - name: text
      dtype: large_string
    - name: model
      dtype: large_string
    - name: volume
      dtype: large_string
    - name: session
      dtype: large_string
    - name: interaction
      dtype: large_string
    - name: participant
      dtype: large_string
    - name: segment_id
      dtype: large_string
    - name: split
      dtype: large_string
    - name: silence_after_ms
      dtype: float64
    - name: prev_same_speaker
      dtype: large_string
    - name: next_same_speaker
      dtype: large_string
    - name: prev_other_speaker
      dtype: large_string
    - name: next_other_speaker
      dtype: large_string
    - name: llm_label
      dtype: large_string
    - name: llm_confidence
      dtype: float64
    - name: llm_rationale
      dtype: large_string
    - name: llm_parse_strategy
      dtype: large_string
    - name: tt_label
      dtype: large_string
    - name: tt_confidence
      dtype: float64
    - name: final_tag
      dtype: large_string
    - name: accepted
      dtype: 'null'
    - name: accept_reason
      dtype: large_string
    - name: reject_reason
      dtype: large_string
  splits:
    - name: train
      num_bytes: 1672719890
      num_examples: 2781985
  download_size: 431528233
  dataset_size: 1672719890
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Seamless Interact - Canary Transcripts

Speech transcription dataset generated by running NVIDIA Canary-Qwen2.5B ASR model on the Seamless Interact conversational speech dataset, segmented by original corpus boundaries.

Dataset Description

This dataset contains 2,781,985 transcribed speech segments from the Seamless Interact corpus. Each segment includes the transcribed text, timing information (offset and duration within the source audio), and metadata identifying the source conversation.

Source

• Audio corpus: Seamless Interact (English conversational speech)
• ASR model: NVIDIA Canary-Qwen2.5B (canary-qwen-2.5b)
• Segmentation: Original corpus segmentation boundaries

Dataset Structure

Column	Type	Description
audio_filename	string	Source WAV filename (e.g., V00_S0030_I00000125_P0045.wav)
offset	float	Start time of the segment within the audio file (seconds)
duration	float	Duration of the segment (seconds)
text	string	Transcribed text
model	string	ASR model used (canary-qwen-2.5b)
volume	string	Volume ID (e.g., V00)
session	string	Session ID (e.g., S0030)
interaction	string	Interaction ID (e.g., I00000125)
participant	string	Participant ID (e.g., P0045)

Example

from datasets import load_dataset

ds = load_dataset("hiraki/seamless-interact-canary-transcripts", split="train")
print(ds[0])
# {'audio_filename': 'V00_S0030_I00000125_P0045.wav',
#  'offset': 0.0,
#  'duration': 3.119,
#  'text': 'James what is your happiest childhood memory',
#  'model': 'canary-qwen-2.5b',
#  'volume': 'V00',
#  'session': 'S0030',
#  'interaction': 'I00000125',
#  'participant': 'P0045'}

Statistics

• Total segments: 2,781,985
• Source files: 93,129 JSONL files
• Format: Parquet (94.3 MB)

Intended Use

This dataset is intended for research on:

• Conversational speech recognition
• Turn-taking prediction and modeling
• Dialogue systems

License

This dataset follows the license of the source Seamless Interact corpus (CC-BY-4.0).