Datasets:
File size: 4,849 Bytes
bf370b5 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 | # Selection rules v1 (provisional_machine_screened). Frozen once subsets are built;
# any change must bump selection_version and keep the old manifests.
selection_version: v1
annotation_status_default: provisional_machine_screened # no human listening yet
normalizer: vitw_en_norm_v1
candidate_generator: cmudict_homophone_nearphone_v2 # alt zipf >= 2.5, max 6 alts
exclude_function_words_from_buckets: true
selector:
model: openai/whisper-large-v3@06f233fe06e710322aca913c1bc4249a0d71fce1
decode: {greedy: true, language: en, prompt: null, fp16: true, word_timestamps: true}
note: >
Whisper has an internal decoder LM; it is NOT a pure acoustic oracle.
HF beam N-best was dropped (transformers 5.4.0 under python -I returned identical beams and OOMed).
acoustic:
# crop = approximate target window (greedy word timestamps; interpolated from
# neighbours if the target was deleted) +/- 0.3 s; decoder sees only the task
# prefix + candidate word (no sentence context).
crop_margin: "min over NON-homophone alts a of [lp_whisper(target|crop) - lp_whisper(alt|crop)]"
crop_margin_all: "same over all alts incl. homophones (diagnostic only; homophone gap = spelling prior)"
homophone_only: "target whose alternatives are all homophones -> acoustic_label homophone_only (crop cannot separate them)"
ambiguous_if: "selector top-1 target wrong OR crop_margin < 1.0"
clear_if: "selector top-1 target correct AND crop_margin >= 3.0"
unknown_if: "otherwise, or window_source == interpolated_too_short"
thresholds: {ambiguous_margin: 1.0, clear_margin: 3.0, max_selector_utt_wer: 0.5}
unreliable_if: "selector utterance WER > 0.5 (hallucination / unintelligible) -> acoustic unknown"
context:
lm: Qwen/Qwen2.5-7B-Instruct (raw text, no chat template) -- nomination only, not ground truth
# margins are natural-log probabilities; min over all alternatives (hardest alternative)
prefix_margin: "min_a [lp(target|prefix) - lp(alt|prefix)]"
prior_margin: "min_a [lp(target|BOS) - lp(alt|BOS)] (context-free word prior)"
prefix_shift: "min_a [(lp(t|prefix)-lp(a|prefix)) - (lp(t|BOS)-lp(a|BOS))]"
full_margin: "min_a [lp(sentence_t) - lp(sentence_a)] # oracle bidirectional diagnostic"
full_shift: "min_a [(lp(sent_t)-lp(sent_a)) - (lp(t|BOS)-lp(a|BOS))]"
high_margin: 2.3 # ~10x odds for the reference over the hardest alternative
high_shift: 1.0 # context must add >= e^1 odds beyond the word prior
low_margin: 0.7 # below this for BOTH prefix and full -> low / undeterminable
availability:
both: prefix high and full high
prefix_available: prefix high, full not high
suffix_only: prefix not high, full high
none: neither high
buckets:
main_challenge: acoustic ambiguous AND context high (any availability)
ambiguous_negative: acoustic ambiguous AND context low
clear_control: acoustic clear AND context high
general_control: acoustic clear AND context low
homophone_semantic: acoustic_label homophone_only AND context high (semantic word choice; NOT evidence that acoustic degradation drives routing)
unassigned: everything else (medium context, alignment unknown, function word) -- kept with reason
baseline_error_enriched: "selector top-1 aligned target word != reference, ignoring surface variants (numbers 7/seven, ordinals, British/American spelling, cannot/can't; surface_equiv_v1)"
inclusion_never_uses: [MoE outputs, MoE routing, MoE accuracy]
partition: {seed: 20261108, dev_frac: 0.2, unit: template group}
decision_log:
- date: 2026-11-09
change: homophone alternatives removed from acoustic crop margin; homophone-only targets moved to separate bucket homophone_semantic
reason: 720 dev homophone targets were labelled acoustically clear only through Whisper spelling prior (plan sec 2.3 requires homophones as a separate class)
data_used: dev distributions only; no MoE outputs
- date: 2026-11-09
change: low_margin kept at 0.7 after dev sweep (0.7/1.0/1.5/2.3 -> low 33/40/55/71 of 3115 dev targets)
reason: low-context targets are rare in this corpus under any threshold; keep conservative definition, report low-context buckets as exploratory
- date: 2026-11-11
change: baseline_error_enriched ignores surface-form variants (surface_equiv_v1)
reason: root review of strict subset found 44 number-format and spelling variants (seven->7, center->centre) counted as selector errors
data_used: strict subset audit only; no MoE outputs
- date: 2026-11-11
change: surface_equiv_v2 (toward/towards, compounds like break room/breakroom); selector utterance WER > 0.5 -> acoustic unknown
reason: root spot-check of 8 random strict-test rows found 2 whole-utterance hallucinations ("Thank you for watching.") and 2 surface variants
data_used: strict subset audit only; no MoE outputs
|