issai/asr-autoadapt-eval-splits

ASR AutoAdapt — Evaluation Splits v2

Rebuilt, cleaned dev / hidden-test splits for two English target domains, used by the protocol-v2 adaptation-agent study (ASR modality). Both domains are derived from public corpora but re-split and cleaned so that the splits are defensible for agent evaluation: conversation-disjoint, distribution- matched, and free of annotation artifacts. All audio is 16 kHz mono PCM-16 WAV; all manifests are JSONL with paths relative to this repository.

domain	source corpus	dev	hidden test
accented	EdAcc (Edinburgh International Accents of English Corpus)	319 utts, 26.8 min, 20 conversations, 13 accents	534 utts, 54.8 min, 23 conversations, 13 accents
medical	PriMock57 (mock primary-care consultations)	396 utts, 30.0 min, 18 consultations	820 utts, 60.0 min, 39 consultations
phone	Switchboard-1 (conversational telephone speech, via hhoangphuoc/switchboard)	370 utts, 30.1 min, 341 calls	770 utts, 60.1 min, 625 calls

Plus the two training-side assets fixed by the protocol (both LibriSpeech):

asset	role	contents
reference/librispeech_clean, reference/librispeech_other	forgetting-gate anchors (dev-clean / dev-other subsets)	500 + 500 utts, audio included
reference_replay/librispeech_train	replay pool the harness mixes into training at the agent's mix_ratio	all 28,539 utts of train-clean-100 (~100 h); audio shipped as 8 tar shards under audio_shards/ with a SHARDS.json checksum index (extracted by the harness's fetch_data.py)

There is deliberately no target-domain training split: in protocol v2 the agent acquires its own training data live, and each run records exactly what it acquired, with source revisions, in its own trace directory. The only training-side asset is the LibriSpeech replay pool above.

Phone (Switchboard) — carried over from study 1, re-verified

Call-disjoint dev/test drawn from hhoangphuoc/switchboard (revision 974526a1…): dev from the train parquets, test from the test parquets, call-id intersection 0. Transcripts arrive pre-normalized upstream (lowercase, no punctuation, contractions expanded). Audio is 16 kHz mono PCM-16 upsampled from native 8 kHz PSTN, so the narrowband acoustic gap is preserved. Re-verified on the rebuilt evaluation stack: the base model's dev/test WER gap is 0.09 pp (17.77 / 17.68). No annotation-artifact rows. Manifests carry call_id, channel, text_raw (upstream text) and split_unit (= call_id).

Base-model calibration (for the record)

Base constants do not transfer across environments, so each study environment re-measures them; these are the values in the study's environment (H100, torch 2.6.0+cu124, transformers 4.44.2, greedy decoding):

model	accented dev	accented test	medical dev
openai/whisper-large-v3-turbo	20.14	19.31	17.16
distil-whisper/distil-large-v3	19.37	19.33	—
omniASR-CTC-300M (fairseq2)	67.59	64.51	—

(WER %, canonical normalizer: lowercase, contraction/number expansion, markup stripped, [a-z0-9 ] only.)

Files

edacc/dev_manifest.jsonl, edacc/test_manifest.jsonl, edacc/audio/*.wav
primock/dev_manifest.jsonl, primock/test_manifest.jsonl,
primock/{dev,test}/audio/*.wav
phone/dev_manifest.jsonl, phone/test_manifest.jsonl, phone/{dev,test}/audio/*.wav
reference/librispeech_{clean,other}/{manifest.jsonl, audio/*.wav}
reference_replay/librispeech_train/manifest.jsonl
reference_replay/librispeech_train/audio_shards/{*.tar, SHARDS.json}
provenance/   split scripts, candidate evaluations, split plan,
              pre-cleanup manifests, artifact-filter note

Audio paths in the LibriSpeech manifests are relative to the parent of the manifest's directory (librispeech_clean/audio/<id>.wav), the original study-1 convention; the harness resolves both conventions.

Manifest fields — accented: id, audio_path, text, duration_s, speaker_id, call_id, accent, l1, gender, split_unit; medical: id, audio_path, text, duration_s, consultation, day, role, split_unit. text is the raw source transcript (EdAcc: upper-case; PriMock57: cased); apply the canonical normalizer to both reference and hypothesis before scoring.

Intended use and hygiene

• dev is visible to adaptation controllers (diagnosis, selection); test is evaluated once per run, after selection freezes.
• Neither eval split may be used as training data; the study's acquisition audit blocks the source corpora (edacc, primock, switchboard, librispeech) and screens acquired items against these transcripts. The replay pool enters training only through the protocol's replay mix, never through acquisition.

Licensing and attribution

This is a private, internal re-packaging. The audio and transcripts remain under their source corpora's licenses and terms: EdAcc (University of Edinburgh; Sanabria et al., 2023), PriMock57 (Babylon Health; Korfiatis et al., 2022), LibriSpeech (Panayotov et al., 2015; CC BY 4.0), and Switchboard-1 (LDC97S62 -- LDC-licensed; hhoangphuoc/switchboard redistributes it without a stated license, so treat the phone domain as academic-use only and never redistribute it outside this private repository). Cite the source corpora in any publication; do not redistribute outside those terms.