routsourav1729/voice-testbed-corpus

Voice Stack Testbed Corpus

Base Hindi two-speaker conversations, plus 1,468 deliberately degraded copies of them (noise, reverb, competing speech, transients, speed changes), used to evaluate a denoise -> VAD -> turn-end detection pipeline component-by-component.

This is test/eval data, not training data: every degraded file traces back to a clean original with a known transcript, so you can score exactly what a denoiser, VAD, or turn-endpointer recovered or lost.

What's in here

data/primary/
  <clip_id>.wav                 37 clean two-speaker Hindi conversations
  transcripts/<clip_id>.json    human transcript: speaker, text, start/end time per segment
  truth/                        derived from the transcripts:
    <clip_id>.vad_speech.npy      0/1 speech label per 32 ms frame
    <clip_id>.turns.json          one turn decision per speaker-change boundary
                                   (TURN_END / TURN_END_OVERLAPPED / NO_FIRE)

out/corpus/
  <scenario>/<clip_id>__<cell>.wav   1,468 degraded copies across 16 scenarios
  manifest.jsonl                     one row per file: scenario, source clip,
                                      and the exact noise/SNR/reverb/etc. it was built with

Where the base clips come from

The 37 base conversations and their transcripts are drawn from Indic DiarBench (Hindi subset). They are included here unmodified except for selection; this corpus does not generate transcripts, it consumes them.

License note: verify Indic DiarBench's own license/terms before relying on these clips outside evaluation use — they are redistributed here as-is under this repo's public visibility, but I have not independently re-verified their redistribution terms beyond that they are published as an open benchmark.

How the degraded copies are made

Each scenario is one recipe: a noise/reverb/level/speed treatment applied to some or all of the 37 base clips, one degraded file per clip per scenario (not one-file-per-parameter-rung — see "one file per category" below).

scenario	what it tests	source of the degradation
00_clean	reference arm	the untouched base clips, resampled/leveled only
01_speed	speaking-rate robustness	WSOLA time-stretch (pitch-preserved), 0.8x/1.2x
10_snr_office	office noise	DEMAND OOFFICE, SNR drawn from {8,5,3} dB
11_noise_class	stationary vs modulated noise	DEMAND (TCAR,PSTATION,PCAFETER), TAU Urban Acoustic Scenes (street_traffic,metro_station,shopping_mall), MUSAN noise, WHAM! ambient
12_competing_speech	crosstalk a denoiser should NOT remove	real Hindi overlap speech from other DiarBench clips, SIR {12,6} dB
14_reverb	room acoustics	RIRs from OpenSLR SLR26 (simulated, small/medium/large) and SLR28 (real recorded)
15_transient_events	impulsive noise (horn, siren, knock, alarm)	ESC-50 event clips, event-SNR {15,12} dB
16_background_bursts	mis-attributed background speech in caller pauses	sparse single-talker DiarBench bursts, duty-cycled, placed in real pauses
20_boardroom	multi-talker reverberant room	DEMAND OMEETING + 2 competing talkers + SLR26 medium reverb
21_boardroom_cued_assent	background "yes/no" during caller pause	one placed DiarBench burst per clip, only where a real >=0.5s pause exists
22_silent_car	stationary low-frequency noise, easy case	DEMAND TCAR
23_railway_metro	announcements + crowd	DEMAND PSTATION, TAU metro_station/metro
24_office_floor	open-plan noise + colleagues	DEMAND OOFFICE + 2 competing talkers
25_street_market	layered urban bed (no single source matches a bazaar)	TAU street_traffic + street_pedestrian + public_square, two pools summed
26_low_volume_noisy	quiet handset in noise	DEMAND OOFFICE + level pulled down to {-30,-24} dBFS
27_worst_case	everything at once, hard end of every ladder	TAU traffic + DEMAND cafeteria + 3 competing talkers + SLR26 large reverb + -24 dBFS

One file per category: where a scenario lists several SNR/event-SNR values, that is the ladder it draws from, not a cartesian product — each clip gets one randomly (seeded) chosen rung, not one file per rung. This keeps the corpus at ~1 file per clip per scenario axis instead of multiplying disk use by the ladder length for a variable nobody scores per-rung.

Noise-pool sources, all standard public corpora:

• DEMAND — Diverse Environments Multichannel Acoustic Noise Database
• MUSAN — music/speech/noise corpus (OpenSLR)
• TAU Urban Acoustic Scenes (DCASE)
• ESC-50 — Environmental Sound Classification
• WHAM! — noisy speech separation corpus (ambient tracks only)
• OpenSLR SLR26 / SLR28 — room impulse responses (simulated + real)

None of those pools are redistributed in this dataset — only the mixed result (base speech + noise, rendered to a single wav) is included, per each pool's own terms.

manifest.jsonl columns

One row per degraded file:

{"scenario": "20_boardroom", "clip_id": "hindi_048", "speaker": "2spk",
 "cell": "...", "file": "20_boardroom/hindi_048__....wav",
 "requested": {"snr_db": 5.0, "sir_db": 12.0, ...}, "tags": ["scenario","boardroom","combined"]}

requested holds the exact parameters that file was rendered with, so a degraded file's condition is always traceable without parsing the filename.

Ground truth format

truth/<clip_id>.turns.json — one entry per speaker-change boundary in the transcript:

• TURN_END — the other speaker was silent long enough for a real VAD stop to have fired; scorable.
• TURN_END_OVERLAPPED — the next speaker started before the first one finished (or the pause never reached the VAD-stop threshold); no inference point exists here in a live system, so it's excluded from turn-detection scoring, not counted as a miss.
• NO_FIRE — a same-speaker pause (not a turn change) long enough that a naive detector might mistake it for one; used to score false turn-ends.

truth/<clip_id>.vad_speech.npy — a 0/1 array, one value per 32 ms frame, derived directly from the transcript segment boundaries.

Corpus limitations (read before quoting numbers)

• The base clips are 95% speech, so VAD false-alarm rate has a small denominator (100 non-speech frames per clip); a detector that always says "speech" would already score close to the top on some metrics.
• Most turn boundaries (roughly 3 in 4) have the other speaker still talking at that instant — no VAD stop occurs there, so no turn-detector inference happens either. Only a minority of boundaries per clip are actually scorable for turn-detection.
• Transcript timings are turn-level, not frame-accurate — good for comparing conditions against each other, not for quoting absolute AUC against a model's published benchmark figures.

Intended use

Evaluating/benchmarking speech-enhancement, VAD, and turn-taking/endpointing models under controlled, labeled degradations — not for training a model to imitate this specific noise mix, and not as a general-purpose noisy-speech training set.