Model Card: DISENT-KWS

Model Description

DISENT-KWS is a speech disentanglement model for robust custom word detection. It uses a shared BC-ResNet-2 encoder with dual heads — a Causal Conformer (phonetic) and ECAPA-TDNN Lite (speaker) — to produce orthogonal latent embeddings. A Dual-Gate Scorer combines keyword and speaker similarity with calibrated weights for joint verification.

  • Total Parameters: 1.806M
  • ONNX Size: 0.60 MB (INT8 quantized)
  • Input: 80-band log Mel filterbank energies, 200 frames (2 s @ 16 kHz)
  • Output: z_phn ∈ ℝ¹⁹² (phonetic embedding), z_spk ∈ ℝ¹⁹² (speaker embedding)

Intended Use

  • Custom keyword spotting with speaker verification for access control, voice assistants, or wake-word systems
  • Real-time streaming inference on CPU (26.4 ms per 2 s window)
  • Few-shot enrollment (5+ utterances) for new users and custom words

Architecture

ComponentTypeParams
Shared EncoderBC-ResNet-233.8K
Temporal BlockMamba SSM / Dilated Conv1D10.3K
Phonetic HeadCausal Conformer (4 heads, kernel 15)1,673K
Speaker HeadECAPA-TDNN Lite (SE ratio 4, scale 4)88.8K
ScorerDual-Gate (w_kw=0.30, w_spk=0.65, EMA α=0.7)

Training Data

DatasetUsageSamples
Google Speech Commands v2Keyword pre-training (35 classes)105K
VoxCeleb1Speaker verification (1,251 speakers)153K
LibriPhraseHard-negative triplet pairs3K triplets
MUSANNoise augmentation (babble, music, environmental)109 hrs

Training Phases

  1. Phase 1: AAM-Softmax pre-training on keyword (GSC) and speaker (VoxCeleb) separately
  2. Phase 2: Joint training with GRL adversarial reversal + CLUB MI minimization + triplet rejection loss
  3. Phase 3a: GE2E speaker head refinement
  4. Phase 3b: Hard-negative GE2E with LibriPhrase confusers

Performance

MetricValue
Keyword EER4.69%
Speaker EER17.86%
Joint EER23.47%
Joint AUC0.8425
CPU Latency26.43 ms (p95: 28.29 ms)
Real-Time Factor (xRT)0.0132
Optimal Threshold (τ)0.2222

Hardware & Runtime

  • Training: NVIDIA GPU with 16 GB+ VRAM (tested on Tesla T4)
  • Inference: CPU-only via ONNX Runtime; no GPU required
  • Memory: ~200 MB RAM at runtime

Known Limitations

  • Requires clean enrollment recordings (SNR ≥ 10 dB recommended)
  • Speaker EER (17.86%) is higher than keyword EER — joint verification mitigates this
  • Mamba SSM fallback to Dilated Conv1D on platforms without CUDA (no performance loss)

License

MIT