Utterly

Dataset Summary

Utterly is a speech dataset derived from pipecat-ai/human_5_all and pipecat-ai/smart-turn-data-v3.1-train. It contains over 7.1k recordings of complete and partial English utterances, each augmented with turn-level annotations, including:

• Verbatim Whisper-generated transcripts
• End-of-turn (EoT) markers
• Speaker identifiers (Coming soon)

The dataset is designed to support research and development of speech and dialogue systems that require joint modeling of speech recognition and conversational turn-taking, such as streaming ASR systems, semantic end-of-turn detection and real-time conversational agents.

---

Source Data

• Base datasets:
  • pipecat-ai/human_5_all
  • pipecat-ai/smart-turn-data-v3.1-train
• Language(s): English
• Modality: Audio (speech; mono-channel; sampled at 16kHz), Text
• Interaction type: Human conversational speech
• Utterances: 7,111
• Speakers: 500+

Dataset splits (e.g., train/validation/test) are not predefined and may be created by downstream users as needed. Deduplication was applied to the underlying audio sources to ensure dataset splits can be made without contamination.

---

Annotation Details

• Transcripts

  • Generated automatically using Whisper Large V3 Turbo.
  • A subset of samples (~200) was manually reviewed and corrected. The transcripts are estimated to have approximately a word error rate (WER) of ~2.8%.

• End-of-Turn markers

  • Human annotations inherited from the base datasets.

• Speaker IDs

  • Coming soon

---

Dataset Structure

A typical data entry includes:

• audio: Path or reference to the audio utterance
• transcript: Text transcription of the utterance
• speaker_id: Identifier for the speaker (Coming soon)
• is_completed: Boolean or categorical flag indicating end-of-turn, i.e. turn completion

---

Usage

In order to load the dataset from the hub, you can use the datasets library:

ds = datasets.load_dataset(
    "ThBel/Utterly",
    split='train',
    streaming=True # (optional)
)

for row in ds:
    # Do something with the data
    print(row['audio']) # or row['is_complete'], row['transcript'], ...

Alternatively you may clone the ThBel/Utterly repository, and load the underlying parquet files using pandas.read_parquet.

---

Intended Use Cases

The Utterly dataset is designed to support a range of speech and dialogue research tasks, including but not limited to:

• Automatic Speech Recognition (ASR) with embedded end-of-turn detection
• Semantic end-of-turn modeling using lexical and acoustic cues
• Turn-taking and floor-control research in conversational AI
• Voice assistants and dialogue systems requiring low-latency response timing

---

Quality Considerations

• End-of-turn annotations involve human judgment and may reflect subjective interpretations of conversational completion.
• Transcription quality may vary depending on audio clarity and source conditions.
• Overlapping speech, interruptions, or disfluencies may introduce ambiguity in turn boundaries.

Users are encouraged to validate performance across multiple evaluation settings.

---

Ethical Considerations

• The dataset consists of recorded human speech and should be used in accordance with the original dataset's licensing and consent terms.
• No additional personally identifying information beyond speaker IDs is introduced.
• Models trained on this dataset should avoid misuse related to surveillance or speaker profiling.

---

Disclaimer and Licensing

Note that Utterly is a derived dataset. I am not the original creator of the source datasets and hold no rights over its content. This dataset is provided as-is for research purposes, and all credit goes to the original authors.

Annotations are released under the BSD-2-Clause license and are intended to be compatible with the licensing terms of the source datasets.

---

Citation

If you use the Utterly dataset in academic or commercial work, please reference the original datasets.