FluidInference/localvqe-coreml

LocalVQE Core ML

Streaming Core ML exports of LocalVQE (weights: LocalAI-io/LocalVQE, Apache-2.0): neural acoustic echo cancellation + noise suppression + dereverberation for 16 kHz speech, a CPU-tuned derivative of DeepVQE (Indenbom et al., Interspeech 2023).

Consumed by FluidAudio (LocalVqeManager / LocalVqeStream); conversion code in FluidInference/mobius models/enhancement/localvqe/coreml.

Files

File	Checkpoint	Params	Samples per call
localvqe-v1.3-4.8M-256ms.mlmodelc	localvqe-v1.3-4.8M.pt	4.8 M	4096 (16 hops)
localvqe-v1.3-4.8M-16ms.mlmodelc	localvqe-v1.3-4.8M.pt	4.8 M	256 (1 hop)
localvqe-v1.2-1.3M-256ms.mlmodelc	localvqe-v1.2-1.3M.pt	1.3 M	4096 (16 hops)
localvqe-v1.2-1.3M-16ms.mlmodelc	localvqe-v1.2-1.3M.pt	1.3 M	256 (1 hop)

All fp32, iOS 17 / macOS 14 minimum deployment target. The two chunk sizes produce identical audio; they trade per-call overhead against latency.

Model I/O

Inputs (Float32): mic [1, N], ref [1, N] (far-end reference — what the loudspeaker played), and 33 in_<state> tensors. Outputs: enhanced [1, N] and the matching out_<state> tensors. Start with all states zero and feed each call's out_* back as the next call's in_*.

The enhanced hop lags the input by 256 samples (16 ms): after consuming input hop k the model emits input hop k−1. The first emitted hop covers t < 0 and can be dropped; feed one hop of zeros at the end to drain. Output level matches the upstream GGML engine (2× the upstream PyTorch reference's overlap-add convention).

Parity / speed

Swift output vs the upstream GGML CLI on the upstream double-talk demo clip: 2.8e-5 max abs diff, 80 dB SNR. Apple M5 Pro, per-call p50 on CPU: v1.3 256 ms 7.1 ms (36× RT), v1.3 16 ms 1.2 ms (14× RT), v1.2 256 ms 4.2 ms (60× RT), v1.2 16 ms 0.7 ms (24× RT).

Citation

Cite the upstream repository (CITATION.cff in localai-org/LocalVQE) and the DeepVQE paper it derives from (Indenbom et al., Interspeech 2023, arXiv:2306.03177).