Talk2NatureEgyptian fruit bat (Rousettus aegyptiacus) / Audiovisual dataset and context-classification review
Fruit-bat calls and context: a rich dataset, not a translation key
Data descriptor and follow-up study years 2017–2023 · Note reviewed 2026-10-12
The 2017 archive describes 293,238 Egyptian fruit-bat WAV files recorded at 250 kHz and 16-bit, plus 91,080 labels derived from synchronized video, including caller, addressee, behavior and social context. The data descriptor reports a post-hoc recheck of 435 annotations, with a 4% estimated context-label error rate (95% CI 2.2–5.8%). The reviewed metadata PDF says annotation FileIDs map to FileInfo.csv; its tables describe recording-date ranges, treatment/channel settings and bats that recur across treatments, and define known, unknown and ambiguous caller/receiver codes. This makes a file/animal grouping audit plausible, but not yet verified against FileInfo rows.
What we reviewed · Targeted primary literature, dataset and rights review
Reviewed the 2017 Scientific Data article's abstract, methods, data records, Tables 3–4, technical validation, usage and rights sections; the 2016 Scientific Reports results and classification methods; the 2023 iScience abstract, dataset description, sequence-HMM methods/results, discussion and data statement; the Mendeley Data version 3 description/license and official version 1/2 and 2/3 comparisons; the Mendeley public file-list API documentation and one anonymous version 3 request (HTTP 401; documentation specifies OAuth2); the NLM catalog entry; and direct public Figshare API metadata for archive item 4556074, annotation item 4555903, FileInfo item 4555897 and Metadata.pdf item 4555900. Acquired only the 3.29 MB annotation table with verified size/MD5 for aggregate-only schema/category-count comparison, plus the 367,257-byte metadata PDF for visual review of all five pages; its source API supplies no MD5, so the receipt's SHA-256 is local provenance only. No audio/video, FileInfo.csv, code or model weights were inspected, and no published model result was reproduced.
Not independently reviewed by a bat bioacoustics or animal-communication specialist.
Study at a glance
- Archive
- 293,238 WAV files (250 kHz, 16-bit); 91,080 video-derived annotations; 435 sampled annotations rechecked, with 4% context-label error estimated.
- 2016 context analysis
- 14,863 calls from seven adult females; four aggressive contexts; 61% balanced accuracy versus 25% chance. Vocalization-level cross-validation is not held-out-animal validation.
- 2023 sequence analysis
- Three adult females; 326 labeled sequences and 2,953 syllables in the HMM; 12,900 n-grams; 66 ± 10 balanced accuracy with random eight-fold cross-validation. Randomized syllable order did not reduce the reported score.
- Figshare item metadata
- The public API declares item-level CC0 1.0 and lists one files210.zip file, 3,049,199,327 bytes. Platform-supplied and computed MD5 fields agree; the payload remains metadata-only and no local checksum was made.
- Annotation-only file audit
- The separate CC0-declared 3,286,405-byte Annotations.csv was fetched with the bounded collector and matched its API MD5. Aggregate-only inspection found 91,080 complete rows, 87,986 distinct FileIDs and exact matches to all context/action category totals in published Table 4. Raw rows are not in the repository or website; this verifies structure and transcription only.
- Metadata PDF review
- All five pages of the 367,257-byte PDF were visually reviewed. The tables map treatment date ranges, social compositions and recording channels; list animal IDs and treatments; and define FileID links plus unknown/ambiguous emitter and addressee codes. This supports a proposed join audit, but does not verify FileInfo.csv contents or any leakage-safe split.
- Rights and status
- The 3.05 GB audiovisual ZIP remains metadata-only; the smaller annotation CSV is stored only under ignored external data and is not admitted to training. Mendeley v3 declares CC BY 4.0, but its versioned file-list request returned HTTP 401 without OAuth2. No media, code or model was acquired or run.
Reading location: Prat et al. 2017, Methods, Data Records, Tables 1–4, Technical Validation, Usage Notes and Rights; Metadata.pdf, visually reviewed all five pages: treatment/date/channel table, subject/treatment table, annotation FileID and signed-identity definitions, and context/action definitions; Prat et al. 2016, Results: Collected data and context identification, and Methods: Classification algorithm and cross validation; Amit and Yovel 2023, Results and Methods: Examining context using HMMs; Mendeley Data version 3 landing page and official versions 1/2 and 2/3 comparisons; official Mendeley API docs and anonymous 401 response; direct Figshare public API metadata for items 4556074, 4555903, 4555897 and 4555900; aggregate-only local audit in research/figshare_bat_annotation_report.json. FileInfo.csv, full supplements and audiovisual archive payload not reviewed.
What this does — and doesn’t — establish
The 2016 analysis reported 61% balanced accuracy for four aggressive contexts versus 25% chance, using 14,863 calls from seven adult females. Its vocalization-level cross-validation does not establish generalization to held-out bats. The 2023 follow-up used sequences from three adult females; the context HMM used 326 sequences (2,953 syllables), formed 12,900 n-grams and reported overall balanced accuracy 66 ± 10 under random eight-fold cross-validation. Shuffling syllable order left the score unchanged. These are predictions of observer-coded context, not evidence that bats or humans interpret a call as a message. The 2023 authors say behavioral experiments are needed to establish biological relevance. The laboratory archive used synchronized video, isolated chambers and specialized 250 kHz ultrasound recording; ordinary phone recordings cannot be assumed to match its bandwidth, calibration or capture conditions. The Figshare public item API declares CC0 1.0 for archive item 4556074 and lists one 3,049,199,327-byte ZIP; its payload and embedded rights notices were not inspected. The separate annotation item 4555903 declares CC0 1.0; its 3,286,405-byte file was locally size/MD5 verified. The aggregate audit found 91,080 rows, 87,986 distinct file IDs, no malformed fields and exact Table 4 count matches. These are structural/transcription checks, not independent validation. The PDF does not prove FileInfo row coverage, uniqueness or actual joins; the 31,574,701-byte FileInfo.csv was not fetched. Its API record has no upstream MD5 and the file exceeds the collector's 5 MiB per-file ceiling. Multiple annotations can belong to one file; ambiguous/unknown identities are present. No raw rows are published. No audio, video, FileInfo contents, encodings, code or weights were inspected or run.
What we can build on
This work offers a strong template for linking sound to independently observed context, auditing annotation uncertainty, and separating acoustic prediction from receiver comprehension. For a new benchmark, split by animal and session, keep context-only and background controls, audit duplicate calls and channels, and reserve receiver-response experiments for a separate hypothesis.
A practical starting point
The annotation table and Metadata.pdf are acquired locally for aggregate/methods review; audiovisual archive, FileInfo.csv and Mendeley deposit remain metadata-only
Available starting point: Video-derived caller, addressee, context and behavior labels; the local aggregate check reproduces the paper's published category totals but does not validate annotation truth
Reuse limits: The direct Figshare API declares CC0 1.0 separately for archive item 4556074 (3.05 GB ZIP, not acquired), annotation item 4555903 (3.29 MB CSV, locally size/MD5 verified), FileInfo item 4555897 (31.57 MB CSV, not acquired) and Metadata.pdf item 4555900 (367 KB; local copy with size and computed SHA-256, no upstream MD5). The annotation table and PDF stay ignored/local; only aggregate and original methodological summaries are published, with no training admission. The paper's CC0 waiver concerns article metadata and its text is separately CC BY 4.0. Mendeley v3 separately declares CC BY 4.0; its file-list API requires OAuth2 and anonymous access returned 401.
Our proposed experiment: Before any model work, freeze a deduplication and held-out-animal/session split, document label definitions and uncertainty, compare acoustic with background/context-only baselines, and state whether outcomes are context association or receiver behavior.
See the related tools and reading paths →
The next reading task
The 3.05 GB audio/video archive remains metadata-only and is not required for this bounded audit. First determine with a qualified bat-bioacoustics reviewer whether acquiring FileInfo.csv is necessary and define a bounded integrity plan for its missing upstream MD5; if justified, inspect only schema, key coverage, duplicate FileIDs, recording times/channels and links to the already audited annotation aggregate. Then test whether animal- and session-grouped splits can be formed without cross-group leakage. Keep annotations out of training unless an independently reviewed protocol and fit-for-purpose benchmark are approved. Mendeley's file-list request needs OAuth2; do not use credentials without explicit authorization. AutoResearch remains NOT READY until all project gates are met.
Go to the sources
- Bat calls, individuals and context ↗
Prat, Taub and Yovel (2016), Scientific Reports 6:39419. Selected primary methods on housing, monitoring and classification tasks reviewed through PMC.
- Egyptian fruit bat audiovisual vocalization data descriptor ↗
Prat et al. (2017), Scientific Data. Reviewed abstract, methods, data records, annotation quality check and rights statement. The descriptor reports 293,238 250 kHz WAV files and 91,080 video-derived annotations; no files were acquired.
- Bat vocal sequences and context classification ↗
Amit and Yovel (2023), iScience. Targeted review of the abstract, reported sequence sample, HMM results, cross-validation description, discussion and data statement. No analysis was reproduced.
- Egyptian fruit bat sequence encodings and example audio, Mendeley Data version 3 ↗
Version 3, published October 20, 2022. The record describes phrase.rar example audio, 326 sequence encodings, syllable encodings and context annotations; it declares CC BY 4.0. The official v2/v3 comparison repeats the same descriptions and license. No file manifest or file contents were inspected or downloaded.
- NLM catalog listing for Egyptian fruit bat Figshare archive item 4556074 ↗
The NLM Dataset Catalog's listing for this Figshare item reports CC0 1.0 and describes Rousettus aegyptiacus social vocalizations. The direct Figshare record could not be fetched during review, so the catalog listing is secondary rights evidence only.
- Egyptian fruit bat Mendeley Data version comparison ↗
Official comparison of Mendeley versions 2 and 3. Both descriptions list phrase.rar example WAVs, 326 sequence encodings, encodings for more than 28,000 syllables and context annotations, and both display CC BY 4.0. The comparison does not verify individual files, upstream audio rights or file contents.
- Egyptian fruit bat Mendeley Data version 1/2 comparison ↗
Official comparison says version 1 contained metadata only; version 2 lists phrase.rar example WAVs, 326 sequence encodings, more than 28,000 syllable encodings and context annotations. The displayed CC BY 4.0 declaration does not establish provenance or rights for individual example audio files.
- Mendeley Data public dataset file-list API documentation ↗
Official API documentation describes GET /datasets/publics/{id}/files with a dataset-version parameter and OAuth2 authorization. An unauthenticated request for dataset mjfv43zgtv version 3 returned HTTP 401; no file manifest or content was retrieved, and no account credentials were used.
- Egyptian fruit bat Figshare item 4556074 public API metadata ↗
The public Figshare API record reports item-level CC0 1.0 and one file, files210.zip, sized 3,049,199,327 bytes. Its supplied and platform-computed MD5 fields match (9c01a0622b2dde0a3d3686ba0263eec3); the payload was not retrieved and the checksum was not verified locally.
- Egyptian fruit bat annotation table, Figshare item 4555903 metadata ↗
The public Figshare API declares item-level CC0 1.0 and lists one Annotations.csv file (3,286,405 bytes) with API-reported MD5 7781a3725d6f2277de74da84b1383ce0. A bounded local retrieval verified size and MD5; the raw file stays in ignored storage and only aggregate audit results are recorded in this project.
- Egyptian fruit bat per-file metadata (FileInfo.csv), Figshare item 4555897 ↗
The official Figshare item API declares CC0 1.0 and lists FileInfo.csv (31,574,701 bytes; file ID 8900695), but md5 is null. The 2017 data descriptor says this file contains exact recording time, recording channel and voiced-segment sample positions. The CSV itself was not downloaded or inspected.
- Egyptian fruit bat subject and annotation metadata (Metadata.pdf), Figshare item 4555900 ↗
The official Figshare item API declares CC0 1.0 and lists Metadata.pdf (367,257 bytes; file ID 8900800), but md5 is null. The 2017 data descriptor says the PDF describes subjects and annotation definitions. Its body was acquired only for local methods review; the receipt records size, TLS provenance and a computed SHA-256, not an upstream checksum.
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.