VITW context-selected subset v1 (strict/prefix; original audio, no added noise)

README.md

@@ -0,0 +1,95 @@
+---
+# Upstream dataset card states no license; verify upstream terms before making this public.
+license: unknown
+language:
+- en
+task_categories:
+- automatic-speech-recognition
+pretty_name: VITW Context-Select (EN)
+tags:
+- speech
+- asr
+- context
+- robustness
+configs:
+- config_name: strict
+  default: true
+  data_files:
+  - split: validation
+    path: data/strict/validation-*.parquet
+  - split: test
+    path: data/strict/test-*.parquet
+- config_name: prefix
+  data_files:
+  - split: validation
+    path: data/prefix/validation-*.parquet
+  - split: test
+    path: data/prefix/test-*.parquet
+---
+
+# VITW Context-Select (EN)
+
+A **selected subset of the original English audio** of
+[`zhifeixie/Voices-in-the-Wild-Bench`](https://huggingface.co/datasets/zhifeixie/Voices-in-the-Wild-Bench)
+(revision `788f5d72c6b0e9091b5c2e432370923b6f9f0660`). **No noise or other processing was added**;
+audio bytes are the upstream WAV files (24 kHz mono PCM16), SHA-256 checked.
+
+Purpose: test whether an ASR system (e.g. an MoE-fusion ASR with an LLM branch) uses **preceding
+sentence context** to recover words that are **acoustically hard to recognise**.
+
+## Configs
+
+| config | rule (per target word; an utterance is kept if ≥1 word qualifies) | validation | test |
+|---|---|---|---|
+| `strict` (default) | acoustically ambiguous vs a near-sounding word **and** the disambiguating context is in the prefix **and** Whisper large-v3 (no prompt) actually misrecognised the word (number-format / spelling variants such as 7/seven, centre/center are not counted as errors) | 85 utts / 0.20 h / 84 groups | 372 utts / 0.90 h / 365 groups |
+| `prefix` | acoustically ambiguous **and** context in the prefix (no Whisper-error requirement) | 283 utts / 0.65 h / 266 groups | 1,076 utts / 2.61 h / 990 groups |
+
+`validation` = dev partition (used for threshold design); `test` = held-out partition. Splits are made at
+**template-group** level (near-duplicate sentences never cross splits), seed 20261108.
+
+```python
+from datasets import load_dataset
+test = load_dataset("ytdata/vitw-context-select", "strict", split="test")
+# only the test files:
+test = load_dataset("ytdata/vitw-context-select", data_files={"test": "data/strict/test-*.parquet"}, split="test")
+```
+
+## How words were selected (machine screening, no human listening yet)
+
+1. **Candidates (text only):** every reference word with a homophone / near-phone alternative from CMUdict
+   (e.g. *leak / league*, *seats / seeds*), alternatives with wordfreq zipf ≥ 2.5. Function words excluded.
+2. **Acoustic ambiguity:** Whisper large-v3 forced-choice on a ±0.3 s crop around the word; ambiguous if the
+   best near-sounding alternative is within e^1 of the target, or Whisper's free decoding got the word wrong.
+   Exact homophones are not counted as acoustic evidence.
+3. **Context:** Qwen2.5-7B-Instruct scores the reference word vs alternatives given the **prefix**
+   (and full sentence); kept when the reference is ≥10× more likely than the hardest alternative and the
+   context adds ≥e^1 beyond the word's prior.
+
+Rule file and all scores are in the build repo (`configs/selection_rules.yaml`, selection v1).
+
+## Columns
+
+`audio`, `sample_id`, `name`, `file_name`, `original_split` (`real_*`/`syn_*` condition), `source_type`,
+`condition`, `mixed_components`, `reference_raw`, `reference_normalized`, `group_id`, `partition`,
+`duration_sec`, `audio_sha256`, `n_targets`, `selector_hypothesis` (Whisper large-v3 greedy output), and
+`targets`: JSON list, one entry per qualifying word with `target_text`, `target_char_span` /
+`target_word_span` (half-open), `target_time_sec` (approximate, from Whisper word timestamps, unverified),
+`alternatives`, `selector_target_word`, `crop_margin`, `context_availability`, `prefix_margin`, `prefix_shift`,
+`full_margin`, `hardest_alternative_prefix`.
+
+## Limitations
+
+- Labels are **provisional (machine-screened)**; nobody has listened to the audio yet.
+- `strict` is enriched for errors of one specific model (Whisper large-v3); it is "hard for that selector",
+  not representative of all ASR or of the full benchmark. Report full-benchmark results separately.
+- Whisper has an internal LM, so "acoustic ambiguity" is a proxy, not ground truth.
+- Text overlap was checked only against the Voices-in-the-Wild-2M English text manifest (no exact or
+  near matches found); overlap with other training data is unknown.
+- Source sentences are short, highly predictable everyday sentences; low-context counterexamples are rare.
+- One truncated upstream clip (`vitw_04834_real_en_far_field`, 0.08 s) is excluded.
+
+## 中文摘要
+
+從 Voices-in-the-Wild-Bench 英文原始音檔（**未加任何噪音**）挑出「目標詞聽起來易混淆、但前文上下文可判斷」
+的句子。`strict` 另要求 Whisper large-v3 實際聽錯該詞（test 372 句）；`prefix` 較寬（test 1,076 句）。
+validation 為 dev（已用於門檻設計），結論請以 test 報告。標籤為機器初篩，尚無人工聽音驗證。

STATS.json

@@ -0,0 +1,50 @@
+{
+  "stats": {
+    "strict/validation": {
+      "utterances": 60,
+      "targets": 78,
+      "groups": 59,
+      "hours": 0.142,
+      "real": 21
+    },
+    "strict/test": {
+      "utterances": 284,
+      "targets": 396,
+      "groups": 277,
+      "hours": 0.693,
+      "real": 63
+    },
+    "prefix/validation": {
+      "utterances": 258,
+      "targets": 387,
+      "groups": 241,
+      "hours": 0.586,
+      "real": 78
+    },
+    "prefix/test": {
+      "utterances": 988,
+      "targets": 1631,
+      "groups": 902,
+      "hours": 2.402,
+      "real": 270
+    }
+  },
+  "files": {
+    "hf_release/vitw_context_select/data/strict/validation-00000-of-00001.parquet": {
+      "sha256": "50752eaf5212d3c233e6dc706d6242a3dc1146045ac146118b18dc4c34ab9247",
+      "bytes": 19554865
+    },
+    "hf_release/vitw_context_select/data/strict/test-00000-of-00001.parquet": {
+      "sha256": "0bd3f56a9ea80c4a97c8aa8de1de5f26be1cc99a588b9640f45b440a49cff2b3",
+      "bytes": 95772293
+    },
+    "hf_release/vitw_context_select/data/prefix/validation-00000-of-00001.parquet": {
+      "sha256": "2817087d441481f54dd0515121806d7933a1e2b1cf3e384c5ca8fa28230689ee",
+      "bytes": 78013803
+    },
+    "hf_release/vitw_context_select/data/prefix/test-00000-of-00001.parquet": {
+      "sha256": "87093b64e9ea2250f7452a58b0c4a329878aecbd3c3dfed644075a6f1593569f",
+      "bytes": 324689761
+    }
+  }
+}

data/prefix/test-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:87093b64e9ea2250f7452a58b0c4a329878aecbd3c3dfed644075a6f1593569f
+size 324689761

data/prefix/validation-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2817087d441481f54dd0515121806d7933a1e2b1cf3e384c5ca8fa28230689ee
+size 78013803

data/strict/test-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0bd3f56a9ea80c4a97c8aa8de1de5f26be1cc99a588b9640f45b440a49cff2b3
+size 95772293

data/strict/validation-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:50752eaf5212d3c233e6dc706d6242a3dc1146045ac146118b18dc4c34ab9247
+size 19554865

selection_rules.yaml

@@ -0,0 +1,73 @@
+# 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