PianoTalk

Project page · Paper · PianoBench

PianoTalk is a multimodal dataset containing over 20,000 hours of aligned speech and piano transcripts from 138,212 instructional YouTube videos. It includes piano MIDI transcriptions, YouTube captions, and speech transcripts generated with NVIDIA Parakeet. We release the raw transcriptions alongside the exact documents used in our paper’s mid-training experiments, which interleave spoken commentary and text-encoded MIDI notes in temporal order. We also provide separate speech and music documents.

Download the raw transcriptions or the mid-training documents.

Dataset layout

Both archives include pianotalk/manifest.json and extract into the layout below.

pianotalk/
├── manifest.json
├── raw/
│   └── <prefix>/<youtube_id>/
│       ├── piano.mid
│       └── captions/
│           ├── youtube.vtt
│           └── parakeet.vtt
└── documents/
    ├── manifest.json
    └── <prefix>/
        ├── <youtube_id>.interleaved.doc.txt
        ├── <youtube_id>.deinterleaved-speech.doc.txt
        ├── <youtube_id>.deinterleaved-midi.doc.txt
        ├── <youtube_id>.speech.doc.txt
        └── <youtube_id>.midi.doc.txt

<prefix> is the first two characters of the YouTube video ID. The dataset manifest records video IDs, channel IDs, titles, source metadata, artifact filenames, and the split for each video. Artifact filenames are relative to raw/<prefix>/<youtube_id>/.

The train, validation, and test splits contain 110,554, 13,837, and 13,821 videos, respectively (approximately 80/10/10). Each YouTube channel belongs to one split. All PianoBench source channels are assigned to test.

Transcriptions

Piano MIDI is generated by automatic transcription of source-separated piano audio (see the paper for details). We include YouTube captions and speech transcriptions generated with NVIDIA Parakeet-TDT-0.6B-v3. Transcriptions and captions may contain errors. Source audio and video are not included.

Documents

The documents use YouTube captions, a 2,000 ms candidate gap, and a 5,000 ms interleave window. Each block occupies one line. We provide five variants:

• interleaved: speech and piano blocks in temporal order.
• deinterleaved-speech: the speech blocks from the interleaved document.
• deinterleaved-midi: the piano blocks from the interleaved document.
• speech: speech split only at silence longer than 2,000 ms.
• midi: piano split only at silence longer than 2,000 ms.

The document manifest records generation settings and per-document measurements.

Piano blocks use self-describing note tables (MIDI-text):

<piano>unit=10ms;advance=5s;notes(midi_pitch,velocity,onset,duration):60,20,0,500;67,20,486,500;</piano>

Each note is represented by its pitch, velocity, onset, and duration, with onsets and durations quantized to 10 ms increments. An advance; marker is added every five seconds, and note onsets are measured from the most recent marker. The text header specifies the time units and note fields.

To regenerate the documents with uv:

uv run generate_documents.py \
  --dataset-dir pianotalk \
  --output-dir regenerated-documents \
  --caption-source youtube \
  --gap-ms 2000 \
  --interleave-window-ms 5000 \
  --workers 16

Use --caption-source parakeet to generate documents from the Parakeet transcripts.

Citation and license

PianoTalk is released under the CC BY-NC-SA 4.0 license. By accessing or using the dataset, you agree to our disclaimer. If you use PianoTalk in your work, please cite our paper:

@article{bradshaw2026pianobenchevaluatingmusicalperception,
      title={PianoBench: Evaluating Musical Perception and Reasoning in Language Models}, 
      author={Louis Bradshaw and Irmak Bukey and Daniel Chenyu Lin and John Thickstun and Chris Donahue},
      year={2026},
      eprint={2611.03839},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2611.03839}, 
}