TeensyVAD: Tiny Voice Activity Detectors for Telephony

All systems evaluated on the same 10 ms frame grid with human labels. Operating points calibrated on held-out AMI dev meetings for every detector (WebRTC aggressiveness swept 0–3; Silero and Energy at stock settings). AUC computed on raw probabilities where available. Test sets: the TEN VAD public set (30 recordings) and 8 AMI SDM meetings (distant mic); neither was used in training or calibration.

Best size per family vs public baselines. Silero ranks best by AUC (87× larger model); teensy-v5 wins the calibrated operating point on AMI and passes FlashVAD's published TEN AUC at a fraction of the size.

*TEN F1 at best-F1 threshold (upper bound) — like-for-like with FlashVAD's published TEN numbers (F1 0.889 / AUC 0.882 / FAR 26.3%). AMI F1 at AMI-dev-calibrated thresholds. Speed = full streaming path, median µs per 20 ms telephony chunk, one core of an Apple M2 Pro. **unoptimized NumPy int8 runtime; artifact 4× smaller than float. v8 models are a research release — v7-GRU96 remains the shipping champion.

teensy-v5 is the current release: an identical architecture and training recipe to v4, with one change — training noise from MUSAN (CC BY 4.0) replaces the NC-licensed share of ESC-50, so the weights are fully commercial-safe. Nothing was traded for it: teensy-v5-80k sets a new family record of 0.8877 TEN AUC (v4-80k 0.880, FlashVAD v0.1 0.882), teensy-v5-40k posts the best room-audio F1 of any generation (AMI F1 0.8853) at half v4-80k's size, and speed is unchanged at 63–64 µs per 20 ms streaming chunk. Thresholds were calibrated on AMI dev; all sizes converged on thr_hi 0.10. Weights, float npz ×4, ONNX float ×4 and ONNX int8 ×4 (22 KB int8 for the 20k) are on the teensy-vad-v5 model card (CC BY 4.0 · © 2026 Pankaj Doharey / Metacritical).

The v1/v2 families (1M training frames) are flat from 20k to 100k parameters — larger nets only overfit. The v3 family (10.7M frames) finds a sweet spot at ~80k. The v4 family, trained on the full ~100 h of LibriSpeech train-clean-100 (28,539 utterances, 37.9M teacher-labelled frames, memmap/float16 plumbing for a 60 GB-scale design matrix), keeps improving to the largest size tested:

(validation frame F1, Silero-teacher labels)

Post-training int8 quantization is essentially free on these models (ΔAUC ≈ 0.000). Quantization-aware training (fake-quant forward pass with straight-through-estimator gradients; a unit test pins simulation ≡ deployed inference) holds that line across the whole size range — int8 matches float32 within ±0.003 real-world AUC at every size, in files 2.7–3.7× smaller. Size-by-size (TEN AUC, float → int8): 20k 0.871→0.876 (int8 beats float), 40k 0.875→0.877, 80k 0.880→0.877, 100k 0.875→0.876. Two standouts: teensy-v4-80k-qat — 88 KB, TEN AUC 0.877, the family's best AMI AUC (0.863), and teensy-v4-qat — 28 KB, TEN AUC 0.876, beats its 87 KB float32 parent. Notably, int8-vs-float is not monotonic in size: quantization sensitivity depends on the learned weight distribution, not parameter count. In pure NumPy, int8 is a size play (no int8 BLAS kernels); real int8 speed lives in ONNX Runtime.

We make one speed claim, precisely: the fastest batched (offline) inference we are aware of among open VADs — 0.064 µs/frame, roughly 15,000× faster than real time. This is architectural, not magic: a stateless MLP over a fixed 10-frame window treats an entire recording as one matrix multiply, while recurrent VADs (Silero, FlashVAD, TEN) must run hop-by-hop through time. In streaming mode, WebRTC's C extension remains the floor (2 µs) and FlashVAD's compiled frontend beats our NumPy one — though our ONNX model alone is faster per single frame than FlashVAD's native figure. At telephony budgets every system here uses < 0.6% of a core; accuracy-per-KB, not speed, is the real axis.

Every number on this page is reproducible: scripts/compare_all.py (tables), scripts/make_charts*.py (charts), scripts/train_all.py (full pipeline, resumable).