Luigi/bilingual-zh-en-multispk-probe

Bilingual zh-TW ⇄ English multi-speaker streaming probe (57 s)

New here? Use v2: bilingual_multispk_v2_45s.wav. v1's sequential turns cannot elicit speaker separation from a windowed streaming ASR (every window is voice-pure, so the baseline tags all four voices Speaker 0 — a probe blind spot, not a deaf model; mechanism documented below). v2 keeps the same 4 voices and adds 3 overlapped pairs with grid-guaranteed shared windows, scoring tags 4/4 with attribution 0.424 on the reference baseline. v1 files remain for provenance. (gold_v2.json, turns_v2.tsv, transcript_v2.txt; details in the v2 section.)

A single deterministic 24 kHz mono clip that interleaves Taiwanese Mandarin and English turns from four different voices, including one turn that code-switches inside the utterance (Mandarin with an English brand name). Built to probe streaming ASR behaviour that single-language, single-speaker test sets cannot reach.

Audio	bilingual_multispk_57s.wav — 24 kHz mono PCM16, 57.16 s, 12 turns
SHA-256	984e60b14cfe78b46d93c75072df20c65722a7ea8a3e7adb758ff43a62afde51
Voices	4 (2 × zh-TW × 4 turns, 2 × English × 2 turns) — every voice appears more than once, so speaker consistency is measurable
Inter-turn gap	300 ms
Languages	zh-TW 26.0 s, en 27.6 s
Code-switch	turn 9: Google 下去。 (Latin token inside a Mandarin sentence)

Files

• bilingual_multispk_57s.wav — the audio.
• gold.json — machine-readable ground truth: per turn turn, speaker (S1…S4), lang, voice (source corpus voice id), start_s, end_s, dur_s, sha (per-clip SHA-256, so a source clip can be traced) and text; plus the schedule order, source filters and the WAV hash.
• turns.tsv — the same table as TSV.
• transcript.txt — human-readable time-aligned transcript with speaker tags.
• bilingual_multispk_v2_45s.wav, gold_v2.json, turns_v2.tsv, transcript_v2.txt — v2 overlap probe (see below): same 4 voices, 3 grid-aligned overlapped pairs + 4 sequential controls, 45 s. Extra per-turn fields: pair, overlap_s, shared_window (the window index both voices provably share), gain_db (pair level-matching, 0.0 = untouched), fresh_clip (clip unseen in v1).

Provenance and licence

All audio and text come from Mozilla Common Voice 17.0, which is released under CC0 1.0 (public domain dedication), so this derived clip carries no attribution requirement — attribution is given anyway:

• zh-TW side: zh-TW test split, filtered to up_votes ≥ 2 and down_votes = 0, sentence length 8–40 characters.
• English side: en test split, same vote filter, sentence length 20–60 characters.
• Speaker identity is client_id from the corpus metadata. Voices were chosen only by clip count (deterministic), never by model output — so the sample cannot leak model-specific selection bias.
• No synthesis, no augmentation, no room impulse: source clips are concatenated with 300 ms of digital silence between turns. MP3 sources are decoded to 24 kHz PCM via FFmpeg (libmp3lame decode only, at 24000).
• The clip is deterministically reproducible from corpus metadata (seed 20260912).

Intended use, and its limits

Use it as a behavioural probe, not as a WER benchmark: it contains only ~97 reference tokens, so confidence intervals are far too wide for an accuracy claim.

Tokenization matters for any scoring: score hybrid tokens (CJK per character, Latin per word), case/punctuation-folded, and script-normalized (Traditional→Simplified via e.g. zhconv). Skipping the script fold is easy to miss and expensive: a model that transcribes correctly in Simplified against this Traditional reference looks ~2.6× worse than it is (26 pp of the apparent error is script mismatch, not misrecognition).

Reference baseline measured with a streaming 1.5 B encoder-decoder ASR (2 threads, CPU phone, RTF 2.40), scored with script folding: WER 14.4 % — English 10.7 %, zh-TW 15.9 % — with zero deletions and zero insertions; the Chinese errors are phonetic/homophone confusions (三→山, 盐→缘, 联→莲).

Speaker labels need a note on how to read them. On this file the baseline emits one Speaker tag for all four gold voices (attribution error 0.567, tags used 1/4) — but that is a property of the windowed streaming protocol, not a dead diarization feature. Root cause, verified by A/B on this exact file: the pipeline carries no speaker state across windows. The encoder cache is reset at every window boundary, the LM's cross-window history is text only, and no speaker-embedding, clustering, or voiceprint machinery exists anywhere in the path — so each chunk's Speaker N: labels are numbered window-locally, and a second tag appears only when one window's acoustic embeddings contain enough contrast between two voices. The 300 ms gaps here align turns to windows so that each window is voice-pure (the previous turn's sliver at the window's leading edge is attenuated by the cold-cache encode), hence all-Speaker 0. Change the within-window contrast with no model change and tags appear: this same file at 100 ms / 1000 ms gaps uses 3 tags; an overlapped two-voice mix splits Speaker 0/1; carrying the encoder state across windows (the only change) moves attribution 0.567 → 0.294 with 2 tags used. What the model cannot do in this protocol is maintain a stable cross-window speaker identity — tags renumber window to window. So do not read the 1-tag run as 'the model hears one speaker'; read it as 'sequential turns are not tracked across windows'.

An earlier version of this card reported WER 30.9 % / zh-TW 39.1 %; that was measured without script folding and overstated the Chinese error rate by ~2.6×. The audio is unchanged (same SHA-256) — only the measurement was corrected.

v2: overlap probe (45 s) — use this for diarization, not v1

v1's sequential turns cannot show speaker separation (see the mechanism above), so v2 adds what v1 lacks: 3 overlapped pairs (2× zh-TW, 1× en — B starts 2.0 s before A ends, mixed at 0.5 gain each) plus 4 sequential controls (300 ms gaps, same voices), 10 turns, 45 s, same corpora/filters/format/seed discipline (seed 650). Each pair's pre-gap is chosen so the first 1.5 s of the overlap provably sits inside one analysis window — asserted at build time and recorded per turn as shared_window, so the test does not depend on lucky alignment. Pair clips under RMS 0.03 are level-matched to 0.05 (recorded as gain_db; controls are byte-identical to source). 3 of 10 clips are unseen in v1 (fresh_clip); the rest reuse v1 voices/clips in the novel overlap arrangement. Honesty notes: overlap makes recognition harder (cocktail-party penalty) — baseline WER 25.9 % on v2 vs 14.4 % on v1 is the harder test, not a regression. And v2 still cannot test stable cross-window identity (tags renumber per window by design, see above).

Baseline (same streaming 1.5 B ASR, script-folded scoring): tags used 4/4 (every gold voice gets a distinct tag), attribution error 0.424 vs v1's 0.567, tag sequence 0,1,2,2,0,3,1,2,3,0. So the model does separate voices that share a window — v1's 1-tag run was the probe's blind spot, not the model's deafness.

Separation is tier-independent (4-arm characterization, same clip): attribution sits at 0.42–0.43 under every arm while numbering renumbers window-locally — lean-p13 is even tag-identical to the default. Full precision and continuous carry do not move attribution either; they move only WER (overlap word recognition), so recognition and separation are independent axes here. Improving attribution needs cross-window identity (model change).

Loading

import json, soundfile as sf
audio, sr = sf.read("bilingual_multispk_57s.wav")   # sr == 24000
gold = json.load(open("gold.json"))
for t in gold["table"]:
    print(t["start_s"], t["end_s"], t["speaker"], t["lang"], t["text"])