routsourav1729/open-noise

Open Noise — background packs and voice stems for audio pipeline testing

Background audio built to be injected under a live signal — a phone call, a voice-agent conversation, any speech recording — so that noise-robustness, voice activity detection, denoising and diarization can be exercised at a signal-to-noise or signal-to-interferer ratio you choose, rather than one baked in.

No file here contains a foreground speaker. Every clip is background only. That is what lets one pack serve every SNR/SIR instead of being fixed at one, and it is what separates this from an ordinary "noisy speech" dataset — you supply the speech.

Applications

• Noise perturbation / data augmentation — mix under clean speech to generate training or evaluation data at controlled SNR.
• Denoiser testing — a spectral or learned denoiser can be scored against a wide, labeled span of noise types (stationary vs. non-stationary, broadband vs. narrowband, impulsive vs. continuous) rather than one arbitrary noise floor.
• Voice activity detection / turn-detection robustness — the voice/ stems are real background speech (not noise), which is the case a denoiser is not supposed to remove and a VAD must not mistake for the foreground speaker.
• Speaker diarization stress-testing — the multispeaker stems are genuine concurrent conversations, useful for testing overlap handling.
• General audio robustness evaluation — 14 real-world acoustic environments, each with three difficulty variants, for building an SNR/scene sweep.

Two independent sets, deliberately not pre-combined:

set	files	each	what it is
noise/	42	10 min	14 acoustic scenes × 3 variants. No voices added.
voice/multispeaker/	5	30 min	Many-speaker chatter, one language per file. Pure voice.
voice/two_speaker/	1	30 min	Exactly two people in conversation. Pure voice.

You pick the room, pick the voices, and set the ratio between them. Every file is 16 kHz mono PCM_16 and seamlessly loopable — the tail is crossfaded into the head, so a random injection offset can run past the end mid-call without a click.

Layout

noise/<scene>/<scene>__a_steady.wav        constant level, nominal event rate
              <scene>__b_eventful.wav      2.5x transient rate, extra bed layer
              <scene>__c_dynamic.wav       level moves over a 14 dB span
voice/multispeaker/<lang>/multispeaker_<lang>_30min.wav
voice/two_speaker/two_speaker_hindi_30min.wav
voice/two_speaker/manual/                  yours — the build never touches it
manifests/<id>.json                        every donor file, offset, event, dropout
metadata.jsonl                             flat index (HuggingFace AudioFolder)
suite.json                                 what the set covers, and what it does not
journeys/                                  earlier design, kept — see the last section

The 14 scenes

Background is drawn per scene — a place, which owns its ambience recordings and the transient sounds that plausibly happen there. One scene per file, held for the whole ten minutes, so a result is attributable to one environment. Variety lives inside the scene: a monsoon bed is cascaded from 40 different rain recordings end to end, a station from three source families summed at once, so nothing loops audibly across ten minutes.

scene	source recordings	beds summed	voices in source
car_interior	DEMAND TCAR · TAU bus	1	no
street_traffic	DEMAND STRAFFIC · TAU street_traffic	1	no
roadwork	DEMAND STRAFFIC + MUSAN	2	no
rain_monsoon	ESC-50 rain · wind	1	no
home_appliance	ESC-50 washing_machine · vacuum_cleaner · MUSAN	1	no
pedestrian_street	TAU street_pedestrian	1	yes
market_square	TAU public_square + street_pedestrian	2	yes
metro_platform	DEMAND PSTATION · TAU metro_station	1	yes (PA)
metro_carriage	TAU metro · bus	1	yes
station_layered	DEMAND PSTATION + TAU public_square + MUSAN	3	yes (PA)
mall	TAU shopping_mall	1	yes
office_desk	DEMAND OOFFICE · MUSAN	1	no
meeting_room	DEMAND OMEETING	1	no
cafeteria	DEMAND PCAFETER · WHAM	1	yes

Two of these exist because nothing else in the set covers them: station_layered sums three beds so the spectrum is dense and non-stationary at once — there is no quiet band left for a spectral estimate to sit in — and roadwork is broadband machinery plus a high transient rate, the class where spectral subtraction handles neither.

"Voices in source" is the honest column. TAU public/pedestrian/mall, DEMAND PSTATION and WHAM cafes are real recordings of occupied places and carry distant speech we did not add and cannot level-control. Nothing in noise/ has voices added; those five scenes inherit them. Per-file flag: incidental_speech_in_source in manifests/.

The three variants

A treatment axis, not three reruns with different seeds — three arbitrary draws of one recipe would give you no reason to prefer any of them. The three also share no audio: each draws its material independently.

• a_steady — the reference bed. Constant level, the scene's nominal transient rate. Use it for a fixed operating point.

• b_eventful — the same room on a bad day: 2.5× the transient rate and one extra bed layer where the scene has a second source to give. Level still constant, so the only thing that changed against a_steady is how much is happening.

• c_dynamic — the room is fixed but your distance from it is not. One continuous bed under a level envelope spanning 14 dB in 25–60 s blocks, joined by 4–10 s ramps, with 1–3 near-silent dropouts of 0.6–1.6 s at −30 dB. An adaptive denoiser, an AGC and a VAD threshold all settle into a constant condition and stop being tested by it; this is the one that does not let them.

  The envelope only ever attenuates — its top is 0 dB, never positive. The span is what matters, and a positive segment would add its own gain on top of the file's crest factor, spending headroom that the peak guard would then take back off the whole file.

c_dynamic is one continuous bed under an envelope, not a sequence of separately rendered blocks. The environment does not change — only how loud it is, as if you walked nearer the road and back — so rendering blocks and crossfading them would put joins into audio with no reason to have any.

The voice stems

file	voices	built from
voice/multispeaker/hindi/	6+	3 concurrent Hindi DiarBench conversations
voice/multispeaker/marathi/	6+	as above, Marathi
voice/multispeaker/nepali/	6+	as above, Nepali
voice/multispeaker/odia/	6+	as above, Odia
voice/multispeaker/urdu/	6+	as above, Urdu
voice/two_speaker/	2	whole two-speaker conversations from the base set

One language per multispeaker file, never mixed. A mixed-language file cannot be un-mixed, and whether a Hindi agent is disturbed by Hindi background specifically is a question worth being able to ask.

The two-speaker stem is stitched longest-conversation-first from data/primary, the 33 two-speaker DiarBench conversations, with 1.5 s crossfades between conversations. Whole conversations first because that keeps real turn-taking — the inter-turn pauses, the overlaps, the backchannels — which is the entire reason to use dialogue rather than concatenated monologue. Joins are between conversations, never inside one.

The base clips are NOT held out of the two-speaker stem, deliberately. The offline scored corpus holds those 33 conversations out of every background so a bed can never contain a voice that is also a test caller. Here the caller is a real person on a live call, so that collision cannot happen, and this is exactly the two-speaker material the stem needs. Do not use this stem underneath the offline scored corpus, where those same clips are the callers. The multispeaker stems keep the hold-out anyway (78–200 min per language is available without it, so it costs nothing).

voice/two_speaker/manual/ is empty and yours. Drop your own conversations there; the build never touches it. Its README states the level convention so anything you add behaves like the generated stem beside it.

Using them

Noise is normalised on whole-file RMS to −30 dBFS (it is continuous); voice on active RMS — speech frames only — to −26 dBFS. Two references, because the two sets have very different crest factors: noise runs to 32 dB (real field recordings carry their own bumps), voice to 22 dB, and one shared reference let the peak guard rather than the target decide 28 of the 42 packs. Anchoring a voice stem on whole-file level would count the talkers' pauses as signal, and the ratio you set would then depend on how much these particular people happened to pause.

noise gain dB = caller_active_dBFS − target_SNR_dB − pack  rms_dbfs
voice gain dB = caller_active_dBFS − target_SIR_dB − stem  active_dbfs

import json, numpy as np, soundfile as sf

meta = {json.loads(l)["id"]: json.loads(l)
        for l in open("metadata.jsonl", encoding="utf-8")}

pack = meta["noise__station_layered__c_dynamic"]
stem = meta["voice__multispeaker__hindi"]
noise, sr = sf.read(pack["file_name"], dtype="float32")
voice, _  = sf.read(stem["file_name"],  dtype="float32")

CALLER_ACTIVE_DBFS = -20.0   # measure on YOUR caller, over speech frames only
TARGET_SNR_DB      = 5.0     # against the noise bed
TARGET_SIR_DB      = 12.0    # against the background talkers

gn = 10 ** ((CALLER_ACTIVE_DBFS - TARGET_SNR_DB - pack["rms_dbfs"])    / 20)
gv = 10 ** ((CALLER_ACTIVE_DBFS - TARGET_SIR_DB - stem["active_dbfs"]) / 20)

n = 60 * sr                                      # however long you need
a = np.random.randint(len(noise)); b = np.random.randint(len(voice))
bg = np.resize(np.roll(noise, -a), n) * gn + np.resize(np.roll(voice, -b), n) * gv

np.resize is safe here precisely because every file is seamlessly loopable.

Read each file's own figure from metadata.jsonl rather than assuming one constant — the peak guard bites differently per file. The ratio you set is a file average; on c_dynamic packs the instantaneous SNR sweeps by design, and window_rms_dbfs_5s in the manifest gives the spread.

Mixing strategy, in short

• Bed pools are screened. musan/noise is a grab bag and many of its files are isolated events with seconds of true silence; cascaded into a continuous bed those became real dead air (3.8 s below −70 dBFS in an early build). A file is usable as a bed only if ≤ 8 % of its frames sit more than 25 dB below its own average and its 2nd-percentile frame stays within 30 dB of it — the fraction test alone passed uniformly-quiet files that still held a half-second hole.
• One-shots are trimmed to the part that actually sounds before levelling. An ESC-50 mouse click is 5 s of near-silence around one 30 ms tick; matching its whole-file RMS drove the transient to full scale. Trimming also makes the manifest timestamps correct.
• Transient crest is capped twice — per event and on the summed event track, both at 20 dB over the bed RMS. The per-event cap alone left overlap uncontrolled: at 2.5× the desk's rate two one-shots land together often, and the global peak guard was then pulling whole ten-minute files down by up to 9 dB to make room for one coincidence.
• Crossfades are equal-power (sqrt ramps). A linear crossfade between two uncorrelated noise beds dips ~3 dB at its midpoint — an audible notch, and a level step a VAD can trip on.
• Normalisation is one global factor per file, never per segment — on c_dynamic the internal level swing is the stimulus.
• Deterministic. Every choice is drawn from a blake2b-keyed generator; the same seed reproduces every file byte-for-byte.

Source datasets

dataset	used for	licence
DEMAND	multi-minute stationary scene beds (channel 1 only)	CC BY-SA 3.0
TAU Urban Acoustic Scenes 2020 Mobile, dev	urban scene beds; 48 kHz downmixed to mono, resampled soxr VHQ	TAU non-commercial research
MUSAN	technical/ambient bed layers — noise subset only	CC BY 4.0
WHAM!	cafe/bar ambience under the canteen scene	CC BY-NC 4.0
ESC-50	scene-gated transients, appliance and monsoon drones	CC BY-NC 3.0
Indic DiarBench (near-field)	every voice stem (hi, mr, ne, or, ur)	see its dataset card

Licence

This is a derivative of the above. The binding terms are the union of theirs, which is effectively non-commercial + share-alike, with TAU's non-commercial research licence on top. Research and evaluation use; not for commercial redistribution. Cite the source corpora, not just this dataset.

Known limitations

• Only noise/ is voice-free of added speech, and five of its scenes inherit distant voices from their source recordings — flagged per file.
• No Indian bazaar or station corpus exists, so market_square and station_layered layer TAU and DEMAND recordings. Closer than any single source; audit before drawing product conclusions from them.
• Reference levels are close but not identical. 36 of the 42 packs land exactly on −30.00 dBFS; six sit lower, down to −35.33. Two causes, both inherent. The source recordings have their own crest — the loudest sample in meeting_room is a scrape inside DEMAND OMEETING at 484 s, nothing we added — and a c_dynamic file with 14 dB of deliberate level variation has a higher peak relative to its own average by construction. Always use the file's own published rms_dbfs. If you would rather gain-match on the loud passages of a c_dynamic pack than on its average, window_rms_dbfs_5s.p95 is in every manifest.
• Mono, 16 kHz. No spatialisation, and no reverb applied on top — the source recordings carry their own rooms.
• These are stimuli, not a labelled benchmark. There is no ground truth, because there is no foreground speaker; the labels belong to whichever call you inject them into.
• Not covered at all: codec, packet loss and jitter (channel effects below the denoiser — they need a trace replayer, not a mixer), and non-Indian background voices.

How this was built

Generated with a custom Python pipeline (not included in this release) that draws exclusively from the public corpora cited below — no proprietary or company-internal audio of any kind. Per file, the pipeline: cascades donor recordings from the relevant corpus end-to-end to fill the duration without audible looping; applies scene-appropriate transient events at calibrated levels; for c_dynamic, applies an attenuation-only level envelope; normalises to one global reference level per file (RMS-based for noise, active-speech-based for voice); and guards peaks against clipping. Every choice (which donor file, which offset, where an event lands) is drawn from a seeded random generator, so the build is fully deterministic and reproducible given the same source corpora.

Every delivered file was re-opened and verified after building: published level matches the manifest, no clipping, the loop seam is continuous, no unintended silent stretches, and correct duration. manifests/*.json records the full provenance per file — every donor recording, offset, and event placement — so any number in metadata.jsonl can be traced back to exactly what produced it.

journeys/ — the earlier design

journeys/ holds five 10-minute tracks that walk between scenes: kerbside, then footpath, then a platform, crossfaded along a plausible-transition graph. They are realistic calls and poor test articles — when a result moves you cannot say which environment moved it — which is why the per-scene packs replaced them. Kept because they are built and some may still want a "realistic whole call" stimulus. They have their own journeys/metadata.jsonl and journeys/suite.json, and are not part of the 48-file set indexed at the root. Regenerate with python -m testbed.corpus.beds --out <dir>.