Datasets:
own-names-bench v1.3.0
Does a speech recognizer spell the names you taught it, and leave alone the ones you didn't?
A public, reproducible benchmark for taught names: invented and uncommon names given to a recognizer ahead of time through its custom-vocabulary feature (keyterms, prompts, phrase boosting, a taught recording). Every system gets the same names in the same order, on the same public-domain audiobook audio (and, from v1.2.0, call-centre, voice-question and meeting audio in per-source suites; from v1.3.0, two sealed held-out test suites), scored by the same code: github.com/virgil-coach/own-names-bench. Built by Virgil (Little Bear Labs, Inc.).
What is in it
| Clips | 1,459 LibriVox pieces (3.5–15 s, 4.16 h) from 12 audiobooks, 12 readers |
| Name mentions | 1,107 mentions of 798 names (499 of invented names, 608 of traditional names), in 1,051 clips |
| Name-free clips | 408 (34 per book), saying no name of the book's list |
| Name lists | each book's own list (18–316 names; 902 entries, 869 distinct spellings, over all books), plus an all-books list per book (its own names first, then every other book's) |
| Chapters | 8 whole LibriVox chapters (1.96 h; 100 min scored), 4 readers, 343 mentions of names from 4 books' lists (28–64 names) |
| Taught recordings | one per name (1,088): the reader saying the name once, cut from another place in the same recording |
Invented-name books: Eric Brighteyes (Haggard), The Jungle Book, Just So Stories (Kipling), Pellucidar, Tarzan of the Apes, The Mad King (Burroughs), The Life and Adventures of Santa Claus, Queen Zixi of Ix (Baum). Traditional-name books: the Iliad (Butler), the Aeneid (Dryden), The Children of Odin (Colum), Legends of King Arthur (Knowles/Malory). Chapters: The Gods of Pegana, The Book of Wonder (two readings), A Dreamer's Tales (Dunsany), The Worm Ouroboros (Eddison).
Suites (v1.2.0)
Four more sources, each in its own folder suites/<name>/ with its own LICENCE.txt and
suite.json (entries, lists, clips, mentions, name-free items). Lists are 40 names per recording:
its taught names, then names taught elsewhere in the suite, never one said anywhere in that
recording. Untaught = the spoken name removed. Every name is taught with one recording of it.
| Suite | Source and licence | Mentions | Name-free | What it adds |
|---|---|---|---|---|
calls |
AppTek Call-Center Dialogues, CC BY-SA 4.0 (this folder is CC BY-SA 4.0) | 740 in 642 turns of 147 calls (210 calls in all), 14 English accents, 120 speakers; 706 scored (34 are mentions of ordinary words the name finder took, listed in not_names) |
417 stretches of 30–60 s (6.0 h) from all 210 calls | conversational speech; names taught by the same speaker (342) or the other party in the call (398); long name-free audio with a realistic list, to measure insertions per minute |
svq |
Google Simple Voice Questions, CC BY 4.0 | 1,424 in 1,272 readings: each name taught from one person's clean reading, tested on four other people (clean, background speech, media noise, traffic noise) | 320 questions | cross-speaker teaching at scale; noise |
ami |
AMI Meeting Corpus, single distant microphone, CC BY 4.0 | none | 119 stretches of 30 s+ (1.1 h) | far-field meeting audio; insertions only |
modulate |
Modulate entity benchmark, MLCommons public-meeting part, CC BY-SA 4.0; audio not re-hosted | 714 in 643 clips of 26 public recordings; an entity taught from its first clip, tested on the others | none | public meetings, real entities; bench fetch cuts the audio from Modulate's repository at the pinned revision and checks it against the PCM hashes in suite.json |
How the names were found:
calls: a turn that overlaps another speaker's turn gives no clip, and a name-free stretch is closed under overlap (it takes in every turn overlapping it) and contains no name, sentence-initial ones included. Six names whose first mention could not be aligned were dropped.calls: AppTek has no entity labels. Names are capitalised words that are not ordinary English (not on a 19k-word list) and not calendar words, titles, interjections or nationalities, joined when adjacent. This is automatic and not hand-checked; in a sample, about nine in ten are real names (the rest are words like "Streamer" or "Theatre"). A hand review of every spelling marked 26 ordinary words asnot_names: they stay on the lists, and scoring leaves them out.svq: the same rule on the question text.modulate: Modulate's own entity spans, dropping offsets that fall inside a word.
Enrollment cuts:
callsandmodulate: located by CTC forced alignment of the reference text with facebook/wav2vec2-base-960h (Apache 2.0), padded by 80 ms (build/align.py).svq: from the word times the dataset ships.
Contamination: Qwen3-ASR and the hosted services may have trained on AMI and on People's Speech (the source of Modulate's meetings); AppTek (new in 2026, evaluation only) and SVQ are unlikely to be in their training data. None of these sources is in the own-names adapter's training data.
Sealed test suites (v1.3.0)
Two held-out suites, each a dataset folder of its own (run one with bench --data <this dataset>/books13 run ...),
sealed before any system was scored on them (every file hashed), and scored once per final system with the benchmark's
own code. They are published byte for byte as sealed and scored: the suite hash recomputes from the files (SHA-256
of the lines <sha256> <path> of every file in the folder, sorted by path), and bench fetch checks it.
| Folder | What | Suite hash | Manifest the scored runs used | Licence |
|---|---|---|---|---|
books13/ |
held-out audiobooks: 20 LibriVox books, 20 readers, 1,100 mentions of 638 names (777 invented, 323 traditional), 1,200 name-free clips (193 min), 2,256 clips of 3.5–15 s, each book's own list (9–176 names) | 3d6d8b65d4810c4f9571e8274bf02659a1bb507e65c7d0c333a48145a41e202f (3,154 files; books13/) |
books13/manifest.json, SHA-256 e02be6a4…8257c8, fingerprint 455a97ea15754481 |
audio and texts public domain (US); annotations CC BY 4.0 |
conv13/ |
held-out conversation: VoxPopuli test and LibriTTS-R test-clean/test-other in eight conditions (as recorded, phone line, far field, street noise, very quiet with names taught up close, a sound-alike on the list, 0.5–2 s snippets, taught by another person), 2,327 mentions, 1,196 name-free pieces (103 min), 3,523 clips, 40-name lists | dbff73e6e5785d0efcf36be582af061b63d59d351cd655ccb6e153ffa850fdd3 (86,867 files; conv13/suites/) |
conv13/manifest.json (the wrapper listing the eight suites), SHA-256 48286d15…, fingerprint 54d0fa007c55db72 |
CC BY 4.0 |
Disclosures:
- books13 was sealed after the final system was frozen (2026-10-27): v7c (
virgil-coach/own-names-qwen3-asrv0.7.0) had already been chosen. No system was run on the suite while it was built. - books13 and LibriSpeech: 11 of its 20 readers are LibriSpeech speakers, and 5 of its books' readings (the suite reader's own reading; 9 LibriSpeech chapters) are in LibriSpeech. None of it is in our adapters' training data; base models and hosted services may have trained on LibriSpeech.
- books13's references are the Gutenberg text placed on the audio by forced alignment (facebook/wav2vec2-large-960h-lv60-self, gated and calibrated on a dev book), not human transcripts. Its names exclude every name of our training data and every reader, LibriVox project, archive.org item, Gutenberg text and work used by this benchmark, its dev sets and our training sets.
- conv13's VoxPopuli is held out by speaker, not by debate: no test speaker is in our conversation training or dev
data, but 265 of the 404 VoxPopuli utterances used come from a plenary debate that also holds other speakers'
training utterances (
session_sharedon every clip). LibriTTS-R is held out by reader and book. No FLEURS and no NOTSOFAR-1 (no recording was held out from our dev data), so the far-field and quiet conditions use simulated rooms. - conv13 lists prefer names never said in our training data; each mention's
groupsays whether its name was (name seen in training) or not. - The cards with every build step, count and check are in the code repository:
build/books_v13/README.md,build/conv_v13/README.md, and the seals inbuild/books_v13/sealed/,build/conv_v13/sealed/. The exclusion lists used to build them are internal and not published.
Files
manifest.json everything the benchmark needs: entries (names, taught recordings),
lists, clips (reference text), mentions (character span in the
reference), name-free clips, chapters (aligned reference segments,
unscored stretches, mentions), name_words, attribution, and the
SHA-256 of every audio file
contamination.json the v1.2.0 check against the v0.3.0 adapter's training data (below)
contamination_v7c.json the v1.3.0 check of every suite against v7c's (v0.7.0) training data (below)
suites/<name>/ the v1.2.0 suites: suite.json, LICENCE.txt, audio/
books13/ the sealed audiobook suite: manifest.json, LICENCE.txt, ATTRIBUTION.csv, audio/clips, audio/enroll
conv13/manifest.json the sealed conversation suite's wrapper, listing its eight suites
conv13/suites/conv13-<condition>/ suite.json, LICENCE.txt, NOISE_ATTRIBUTION.csv, audio/clips, audio/enroll
audio/clips/<book>/<piece>.flac
audio/enroll/<book>/<name>.flac, audio/enroll/chapters/<book>/<term>.flac
audio/chapters/<chapter>.flac
Audio is 16 kHz mono FLAC, exactly as cut from the LibriVox MP3s (no re-encoding after the cut).
Use it through the benchmark code (bench fetch checks every file against the manifest).
How it was made
- Source:
lore-audiobook-names(test split) andlore-eval-sets(long), our own internal datasets: LibriVox chapters cut at pauses, recognized with Parakeet TDT 0.6B v3, matched to the Project Gutenberg text (pieces kept when at least 75% of words agree) and force-aligned with torchaudio MMS_FA. Names are words the book always capitalises and no word list knows; test name lists were read and corrected by hand. The script isbuild/build_dataset.pyin the code repo. - Mentions: per book and name, up to 3 mentions for invented-name books and 1 for traditional-name books, each from a different piece, sampled with seed 20261013. Name variants (inflected forms) are excluded.
- Taught recordings: each name's first enrollment cut (said by the reader who says it most). No clip is a piece an enrollment cut was taken from. In the chapters, the places the taught recordings were cut from, and unaligned stretches, are unscored holes.
- Name-free clips: 4–15 s pieces with no marked name and no word sequence of any name of the book.
- The Book of Wonder is read by two readers sharing one name list; each name was enrolled once, so 60 of the 343 chapter mentions are taught in the other reader's voice (flagged in the manifest).
Contamination check
v1.3.0, v7c (contamination_v7c.json, build/contamination_v13.py): every suite against the training data of
virgil-coach/own-names-qwen3-asr v0.7.0 (v7c), from the item-by-item attribution in that model's attribution/
folder (which includes v0.3.0's items). PASS, 0 failures: no reader or speaker, LibriVox recording,
training-audiobook text, LibriTTS-R utterance, speaker or chapter, FSD50K clip or training corpus is shared with the
v1.2.0 audiobooks, the four v1.2.0 suites, books13 or conv13. Reported: 125 of conv13's 199 VoxPopuli plenary sessions
also hold other speakers' training utterances (held out by speaker, above), and listed names that are also in a
training name list (v1.2.0 audiobooks 1 of 1,055; calls 2, svq 1, ami 2, modulate 2; books13 0 of 894; conv13 11 of
1,019). Not checkable from the attribution: YODAS voices, VoxPopuli speakers (checked by conv13's build against the
conversation train/dev data: 0), and names said in conversation training utterances (recorded for conv13 only).
v1.2.0, v0.3.0:
The system this benchmark was first built to test is Virgil's own-names adapter for Qwen3-ASR
(own-names-qwen3-asr-lora v0.3.0). Every benchmark book, reader and name was checked against
everything that adapter trained on (contamination.json, build/contamination.py):
- Books: none of the 17 benchmark books is among the 47 LibriVox training books.
- Readers: none of the 16 benchmark readers (clips, chapters and taught recordings) reads a training audiobook, and none is among the 71 LibriTTS speakers whose voices were cloned for synthetic training speech or whose real recordings were used (matched by name through LibriSpeech's SPEAKERS.TXT). The internal test split had five more books read by Mark Nelson, who is LibriTTS speaker 251 and in that training data; all five were dropped.
- Names: no benchmark name is among the 3,000 generated training names or in the name list of a training book. Five Norse names (Allfather, Midgard, Mimir, Niflheim, Ymir) are also listed for a training book and were dropped.
- Not checkable: the adapter also trained on YODAS YouTube clips, which carry no speaker or channel; a benchmark reader appearing in a training video cannot be ruled out.
- Base models: every system's base model may have heard LibriVox. 11 of the 16 benchmark readers are LibriSpeech speakers (training data of many recognizers), and hosted services do not publish their training data. This affects every system, and the texts are classics that language models know.
Versions
- v1.3.0 (2026-10-27): adds the sealed suites
books13/andconv13/andcontamination_v7c.json. The v1.2.0 files are unchanged, and so is the data format (version1.2.0 in every manifest); benchmark code v1.3.0 reads it. - v1.2.0 (2026-10-14): adds the four suites above. The audiobook data, lists, references and mentions are unchanged, and so are every system's audiobook results.
- v1.1.0 (2026-10-14): same audio, lists, references and mentions as v1.0.0 (the manifest's content fingerprint is unchanged; only its version label and change note differ). The scoring code changed: untaught passes drop every entry spelled like the spoken name (65 all-books passes had kept another book's entry of the same spelling), names are matched one to one, intervals are clustered bootstraps, and 65 mentions whose name another book lists in a different reader's voice get a cross-reader teaching pass.
- v1.0.0 (2026-10-13): first release.
Licence and attribution
- Audio: LibriVox recordings, in the public domain in the USA (not guaranteed elsewhere).
Thank you to the readers: Peter Dann, George Allen, Elizabeth Klett, Thomas Rose, Brett W.
Downey, Meredith Hughes, Ralph Snelson, Tim Bulkeley, Mark F. Smith, Delmar H Dolbier, Kristin
LeMoine, Phil Chenevert (clips); Jason Mills, David Mack, Greg Elmensdorp, Sandra Cullum
(chapters). The LibriVox item and source recording of every clip and chapter are in
manifest.json(attribution, andsourceper clip). - Text: from the Project Gutenberg texts of the same books (public domain in the USA), with every Project Gutenberg header and notice removed.
- Annotations (lists, spans, alignment, manifest): Apache 2.0, © 2026 Little Bear Labs, Inc.
books13/: LibriVox recordings, dedicated to the public domain by their readers (US), and the Project Gutenberg texts of the same books (public domain in the USA; not guaranteed elsewhere); our annotations CC BY 4.0, © 2026 Little Bear Labs, Inc. (books13/LICENCE.txt). Thank you to the readers: skoval, klbonds, meyerli, Gabriela Cowan, Scott Robbins, Jacquerie, Lynda Marie Neilson, Katharina21, tjfria, Carol Box, MrsLSid, Mike Harris, MaryAnn, Deon Gines, Kenneth Sergeant Gaghan, Lynette Caulkins, Warren Kati, Hox, Michael Scherer, Andy Sames. Every book's LibriVox page, archive.org item, Gutenberg id and public-domain basis is inbooks13/ATTRIBUTION.csv.conv13/: CC BY 4.0 (LICENCE.txtin each condition). Audio from VoxPopuli (Changhan Wang et al., VoxPopuli, ACL 2021), test split, European Parliament recordings, CC0 1.0 (source: European Parliament); LibriTTS-R (Yuma Koizumi et al., LibriTTS-R, Interspeech 2023; from LibriTTS, Heiga Zen et al., 2019, from LibriVox), test-clean and test-other, CC BY 4.0; background noise from FSD50K (Eduardo Fonseca et al., FSD50K, IEEE/ACM TASLP 2022), clips under CC0 1.0 or CC BY 3.0, each credited by title, Freesound uploader, licence and URL in the condition'sNOISE_ATTRIBUTION.csv. The audio was resampled, mixed, filtered and reverberated (simulated rooms: pyroomacoustics, MIT).- Suites: each folder
suites/<name>/carries its source's licence in itsLICENCE.txt, for the audio and for our annotations of it (suite.json):calls(AppTek) andmodulate(Modulate) CC BY-SA 4.0,svq(Google) andami(AMI) CC BY 4.0. Themodulatefolder holds only our annotations; its audio stays in Modulate's repository and is read from there at a pinned revision.
Citation
own-names-bench v1.3.0 (2026). Little Bear Labs, Inc. https://github.com/virgil-coach/own-names-bench
- Downloads last month
- 166