Multiple mammals and birds; monk parakeet subset / Cross-species bioacoustic methods benchmark

Cross-species pitch benchmark helps estimate vocal pitch, not meaning

Peer-reviewed article and Dryad dataset 2025 · Note reviewed 2026-10-24

The article reports 250,670 calls from 14 taxa with annotated fundamental-frequency (F0) contours. The Dryad release lists waveforms and annotations for 13 taxa; it also includes F0 predictions for bottlenose dolphins whose waveforms could not be published. The listed release totals 27.86 GB. Rodents contribute 224,705/250,670 calls (about 89.6%), while monk parakeets contribute 233. In the paper’s target-trained CREPE comparison, monk-parakeet pitch accuracy is 0.30 under that study’s half-semitone metric. This is an acoustic-feature estimation score, not a communication or meaning score. The 2025 paper cites Smith-Vidaurre et al. (2021, 2023) papers and deposits as sources for the 233 semi-automatically annotated monk-parakeet calls. The 2021 Figshare record lists 12 CSV/KML/README files and no waveform. The 2023 Figshare v2 inventory lists a 440,873-byte selection table and a 31,079,669-byte WAV archive; its README says their sound.file names correspond across those files and reports 1,596 calls in the table, 1,582 used in the paper. This offers a possible metadata-only source-file comparison, but the exact mapping of the F0 paper's 233 calls to the 2023 files remains unverified. The Dryad API v9 manifest lists 15 files totaling 27,863,357,661 bytes and identifies monk_parakeets.zip as 5,076,570 bytes with a provider-reported SHA-256. This is inventory metadata only: the archive and its file_list.txt have not been downloaded or inspected, and the checksum has not been recomputed locally.

What this does — and doesn’t — establish

The benchmark targets a single F0 contour and discards overlapping sections; it does not label call category, behavioral context, addressee, receiver response or meaning. It is highly imbalanced by taxon, combines manual, semi-automatic and automatic annotation procedures, and model performance varies with signal characteristics. The 14-taxa article count and 13-taxa waveform deposit differ because dolphin predictions are present without the dolphin waveforms. The paper, data, code and weights have separate reuse terms. The exact source-recording provenance and independent-unit split structure were not reconciled in this review. The pinned 75-entry GitHub tree has no LICENSE/COPYING file; code and checkpoint terms remain unresolved. The 2025 article's main text does not give a call-to-file crosswalk. No archive or selection-table payload was downloaded, rows and KML contents were not inspected, and provider checksums were not recomputed locally. The 2023 README's file-name relationship is a lead, not a verified match to the F0 subset; source-recording provenance and location sensitivity remain open. The official Dryad file-download API returned an authorization requirement and the public file-stream route returned HTTP 403 in this environment; no archive member names or audio were inspected.

What we can build on

This is a useful candidate for testing whether an acoustic pipeline can estimate a measurable feature such as pitch on non-human calls. It could support a carefully bounded feature benchmark after rights, provenance and grouping checks. It cannot train or validate a system that claims to understand what an animal said. Keep the 233-call monk-parakeet subset distinct from much larger taxa and report per-taxon results rather than a pooled headline score.

See the related tools and reading paths →

The next reading task

The API manifest fixes the current Dryad version and provider checksum, but retrieving the archive through the documented route requires bearer authentication and the public file-stream route returned HTTP 403 in this environment. Do not bypass collector limits or use an unapproved credential. If a documented public route to the metadata-only file_list becomes available, compare it with the 2023 selection-table sound.file values without inspecting audio or location fields; otherwise keep the 233-call subset metadata-only. Any later file-level work must confirm individual/session/source-study keys and location sensitivity, freeze a group-held-out pitch-tracking evaluation, and obtain independent bioacoustics and ML/statistical review before acquisition or model execution.

Go to the sources

1. Bioacoustic fundamental frequency estimation: a cross-species dataset and deep learning baseline ↗

   Best et al. report a cross-species benchmark for estimating single fundamental-frequency contours from non-human vocalizations. It evaluates pitch estimation, not call meaning.

2. F0 estimation for bioacoustics: Dryad dataset ↗

   Dryad record for the F0 dataset: more than 250,000 calls, waveforms and annotated contours for 13 taxa, plus dolphin predictions without dolphin waveforms. The official API returned dataset version 9 with 15 files and a provider checksum for monk_parakeets.zip.

3. Dryad reuse guidance and CC0 policy ↗

   Dryad states that its datasets are published under CC0 and recommends scholarly attribution. This policy source does not license separately published papers, code, model weights or assets outside the deposited dataset.

4. Bioacoustic F0 estimation code and model repository at pinned revision ↗

   Pinned GitHub tree for the paper-linked implementation. The reviewed 75-entry tree contains code and model assets but no LICENSE or COPYING file; code and weights are separate components with unresolved terms.

5. Individual vocal signatures show reduced complexity following invasion ↗

   Peer-reviewed 2021 monk-parakeet study of contact-call frequency-modulation structure and individual identity content across native and introduced populations.

6. Simpler signatures post-invasion dataset (Figshare v1) ↗

   Official Figshare v1 API record declares item-level CC BY 4.0 and lists 12 files totaling 11,338,100 bytes: CSV, KML and README files. The inventory lists no waveform file.

7. Individual identity information persists in learned calls of introduced parrot populations ↗

   Peer-reviewed PLOS Computational Biology study comparing acoustic identity patterns in native- and introduced-range monk parakeets, with yellow-naped amazons as a regional-dialect comparator.

8. Smith-Vidaurre et al. identity-encoding calls and selection table (Figshare v2) ↗

   Official Figshare dataset record for the paper. Its 22-file inventory totals 187,820,138 bytes under an item-level CC BY 4.0 declaration. The README describes a 1,596-row selection table and a WAV archive with matching sound.file names; 1,582 calls were used in the paper. No table or audio was retrieved.

Original editorial note by Talk2Nature contributors. AI-assisted, source-grounded synthesis; not a peer review or an independent replication. Linked works retain their own rights.