binhpham/wakeword-bench-hey-livekit

Wake word benchmark: "hey livekit"

Generated with livekit-wakeword-bench (format v1). Speech is synthesized through LiveKit Inference with many providers and voices (key phrases steered with phonemes), then mixed with real background noise at controlled SNR buckets, optional room reverb and band-limiting. Every clip carries the metadata needed to slice results (noise level/type, reverb, voice, accent, provider, phrase, timing).

Test voices and test noise/RIR sources are disjoint from train/validation ones, so the test split measures generalisation to unseen speakers and environments.

Usage

from livekit_wakeword_bench import WakeWordData, benchmark

data = WakeWordData("binhpham/wakeword-bench-hey-livekit", split="test")
noisy_misses = data.filter(category="positive", noise_bucket=["very_noisy", "extreme"])

result = benchmark("livekit:hey_livekit.onnx", "binhpham/wakeword-bench-hey-livekit")
print(result.summary()["at_threshold"])

Plain datasets works too: load_dataset("binhpham/wakeword-bench-hey-livekit", split="test").

Splits

split	rows	hours
test	7873	12.26
train	23000	23.30
validation	2496	3.46

test

• category: positive: 3000, adversarial: 3000, speech: 1000, real_speech: 573, noise: 300
• noise_bucket: noisy: 1317, extreme: 1281, very_noisy: 1274, clean: 1261, quiet: 1239, moderate: 1201, noise_only: 300
• noise_category: : 1261, babble/restaurant_bar: 359, public/station: 355, street/traffic: 340, public/square: 331, transport/car: 243, office/meeting_room: 226, office/office: 225, transport/bus: 224, office/hallway: 208, transport/metro: 204, babble/cafeteria: 162
• provider: inworld: 1773, rime: 1771, xai: 1770, cartesia: 1686, : 300, earnings22: 209, peoples_speech: 184, ami: 180
• accent: US: 3901, UK: 966, unknown: 885, : 873, IN: 599, MX: 223, DO: 182, DE: 127, AU: 117
• reverb: False: 4104, True: 3769

train

• category: positive: 10000, adversarial: 10000, speech: 3000
• noise_bucket: clean: 23000
• noise_category: : 23000
• provider: rime: 5795, inworld: 5790, cartesia: 5737, xai: 5678
• accent: US: 14478, unknown: 5279, UK: 1378, AU: 810, MX: 341, IN: 333, FR: 199, ME: 182
• reverb: False: 23000

validation

• category: positive: 1000, adversarial: 1000, speech: 300, noise: 100, real_speech: 96
• noise_bucket: noisy: 448, quiet: 418, moderate: 403, extreme: 390, very_noisy: 380, clean: 357, noise_only: 100
• noise_category: mixed/audioset: 1096, mixed/musan_sound-bible: 523, mixed/musan_free-sound: 520, : 357
• provider: rime: 608, xai: 586, cartesia: 560, inworld: 546, : 100, librispeech: 96
• accent: US: 1456, unknown: 509, : 196, UK: 151, AU: 76, MX: 43, IN: 35, FR: 15, ME: 15
• reverb: False: 1327, True: 1169

Audio sources

pool	kind	name	description
train	noise	musan_noise	MUSAN noise (CC BY 4.0). What livekit-wakeword trains with.
train	noise	dns_noise	DNS Challenge noise, AudioSet part only (CC BY 4.0)
train	rir	mit_rir	MIT environmental impulse responses
train	speech	librispeech	LibriSpeech test-clean (CC BY 4.0), read speech for validation FPPH
test	noise	demand	DEMAND multichannel environmental noise, 1 channel per scene (CC BY 4.0)
test	noise	wham	WHAM! restaurant/bar babble, test split (CC BY-NC 4.0)
test	noise	esc50	ESC-50 environmental sounds, folds 1-2 (CC BY-NC)
test	noise	gtzan	GTZAN music, 10 genres (research use)
test	rir	slr28_rir	OpenSLR 28: real RIRs (RWCP/REVERB/AIR) + simulated rooms (Apache 2.0)
test	rir	treble10_rir	Treble10 furnished-room RIRs (CC BY-NC-SA 4.0)
test	speech	ami	AMI meeting corpus, far-field single mic (CC BY 4.0)
test	speech	earnings22	Earnings-22 accented earnings calls (CC BY-SA 4.0)
test	speech	peoples_speech	People's Speech test shard (CC BY)

Speech: LiveKit Inference TTS (Cartesia Sonic 3, Inworld TTS-2, xAI, Rime Mist v3); wake words and confusables are sent as phonemes for exact pronunciation. Some noise/RIR sources are non-commercial; check them before commercial use.

Columns

audio (16 kHz FLAC) plus: id, split, label, category, wake_word, text, text_form, phrase, adversarial_distance, provider, model, voice_id, voice_name, gender, accent, held_out_voice, speed, noise_bucket, snr_db, noise_source, noise_type, noise_category, noise_file, reverb, rir_source, rir_file, rir_category, interferer, sir_db, noise2_category, eq, bandlimit, clipping, codec, augmentations, level_dbfs, onset_s, offset_s, phrase_onset_s, phrase_offset_s, duration_s, tts_key, phonemes, tts_input

Key ones: label (1 = contains the wake word), category (positive / adversarial / speech / real_speech / noise), noise_bucket (clean / quiet / moderate / noisy / very_noisy / extreme / noise_only), snr_db, phrase_onset_s/phrase_offset_s (exact wake word timing, used for detection latency), held_out_voice.

Generation config

name: hey_livekit
wake_word: hey livekit
positive_texts:
- Hey LiveKit
- Hey, LiveKit!
- Hey LiveKit?
- Hey LiveKit.
- Hey, LiveKit
pronunciations:
  livekit: L AY1 V | K IH2 T
hard_negatives:
- hey live kid
- hey lift kit
- hey liv kit
- hey lifekit
- a live kit
- hey, love it
- hey leave it
p_carrier: 0.35
language: en
seed: 1234
sample_rate: 16000
cache_dir: .cache/lkwb
sources: []
tts:
  models: []
  plan: scale
  concurrency: null
  sample_rate: 24000
  test_voice_fraction: 0.3
  speed:
  - 0.85
  - 1.2
  retries: 3
  phonemes: true
augment:
  snr_buckets:
    quiet:
    - 20.0
    - 30.0
    moderate:
    - 10.0
    - 20.0
    noisy:
    - 5.0
    - 10.0
    very_noisy:
    - 0.0
    - 5.0
    extreme:
    - -5.0
    - 0.0
  snr_weights: {}
  clean_weight: 1.0
  p_reverb: 0.5
  p_interferer: 0.2
  sir_db:
  - 0.0
  - 20.0
  p_second_noise: 0.2
  p_eq: 0.4
  p_bandlimit: 0.2
  level_dbfs:
  - -30.0
  - -3.0
  p_clip: 0.05
  p_codec: 0.25
  lead_s:
  - 0.3
  - 1.5
  tail_s:
  - 0.4
  - 1.2
  segment_s:
  - 8.0
  - 30.0
splits:
  train:
    n_positive: 10000
    n_adversarial: 10000
    n_speech: 3000
    real_speech_minutes: 0.0
    n_noise: 0
    augment: false
  validation:
    n_positive: 1000
    n_adversarial: 1000
    n_speech: 300
    real_speech_minutes: 30.0
    n_noise: 100
    augment: true
  test:
    n_positive: 3000
    n_adversarial: 3000
    n_speech: 1000
    real_speech_minutes: 180.0
    n_noise: 300
    augment: true