lars1234/pitch-benchmark

Pitch Benchmark

Evaluation and training data for the pitch tracker benchmark at https://github.com/lars76/pitch-benchmark. Two archives:

File	Size	Contents
eval.tar	382 MB	ten evaluation corpora, test and calibration clips, nine renderings each, reference labels, manifests
train.tar	4.5 GB	five training corpora with labels and the augmentation pools

Checksums are in SHA256SUMS.

Use

hf download lars1234/pitch-benchmark --repo-type dataset --local-dir dataset
tar -xf dataset/eval.tar  -C dataset
tar -xf dataset/train.tar -C dataset

This gives dataset/eval and dataset/train. Pass them to the benchmark and the training code with --dataset dataset/eval and --dataset dataset/train. The layout, the manifests (dataset.json, clips.json, renders.json) and the source corpora are described in the repository.

License

The audio and labels are derived from the corpora below. Copyright in the underlying recordings remains with their original holders, who distribute them for research. This compilation is released for research use under CC BY-NC-SA 4.0 and grants no rights beyond those of the sources. Please cite the sources whose clips you use.

Evaluation corpora (eval.tar)

• KEELE - 10 speakers reading the North Wind passage with a laryngograph (Plante et al., Eurospeech 1995)
• FDA - Bagshaw/CSTR, 50 sentences x 2 speakers, studio 20 kHz with a laryngograph (Bagshaw et al., Eurospeech 1993)
• APLAWD - 151 utterances x 10 British-RP speakers, speech + laryngograph (Lindsey, Breen and Nevard, UCL 1987; Brookes's APLAWDW repackaging)
• AVID - Aalto Vocal Intensity Database: 50 speakers, calibrated ~15 min recordings, speech + EGG (Alku et al., Speech Communication 2024)
• OSF Glottis - Harvard sentences with EGG and intraoral pressure (Rosen, Veillette and Nusbaum, 2024)
• Saarbruecken Voice Database - German connected-speech phrases with EGG, healthy-control subset only (Pützer and Barry)
• SpeechSynth - synthetic Mandarin speech with a known pitch contour, generated from the LightSpeech checkpoint of fastspeech2-clean; voices trained on AISHELL-3 (Shi et al., 2020) and Data Baker CSMSC
• Vocadito - Solo vocal recordings (Bittner et al., 2021)
• URMP - Classical chamber pieces with manually corrected per-track f0 (Li et al., IEEE TMM 2019)
• Bach10-mf0-synth - Resynthesized Bach10 with exact f0 (Salamon et al., ISMIR 2017; Bach10: Duan et al., IEEE TASLP 2010)

Background scenes and rooms rendered into the evaluation clips use DEMAND ambience (Thiemann, Ito and Vincent, 2013), AISHELL-3 speech, MIR-1K accompaniment (Hsu and Jang, IEEE TASLP 2010), and measured impulse responses from RIRS_NOISES (Ko et al., ICASSP 2017) and OpenAIR (Audiolab, University of York).

Training corpora (train.tar)

• MDB-stem-synth - Resynthesized MedleyDB stems with exact f0 (Salamon et al., ISMIR 2017; MedleyDB, Bittner et al., ISMIR 2014)
• NSynth - Acoustic instrument notes labelled by nominal MIDI pitch (Engel et al., 2017)
• PTDB-TUG - Read speech + laryngograph, 20 speakers (Pirker et al., Interspeech 2011)
• MOCHA-TIMIT - 8 speakers x 460 TIMIT sentences + laryngograph (Wrench, Queen Margaret University College, 1999)
• CMU Arctic - Read speech + EGG channel (Kominek and Black, SSW 2004)

Augmentation pools use LibriSpeech speech (Panayotov et al., ICASSP 2015), MUSDB18 stems (Rafii et al., 2017), TAU Urban Acoustic Scenes 2019 (Heittola, Mesaros and Virtanen, 2019), and simulated impulse responses from RIRS_NOISES. Coloured noise is generated by the preparation code.