FRCRN_SE_16K for Tenstorrent Tensix

FRCRN_SE_16K with complex convolution weights written as real block matrices, batch normalization folded into the convolutions, and weights stored in bf16. The tensor layout is suitable for TT-NN implementations. No additional training.

Formats

model.safetensors contains the two UNets' real tensors. Every stored tensor is bf16. config.json records each convolution's stride and padding and each FSMN memory's axis and order.

A complex convolution uses the real block matrix [[Wr, -Wi], [Wi, Wr]]. The following batch normalization is folded into its output channels. Decoder skip-input channels follow the order recorded in config.json. Tensor names begin with u1 or u2, followed by the module name.

The input is a normalized complex spectrum [1, 2, 321, 51] for 16 kHz audio, a 640-point FFT and 320-sample hop. The output is the complex mask tanh(u1) + tanh(u2) in the same real/imaginary layout. The consuming implementation supplies framing, normalization, mask application and synthesis.

Use

Read one tensor with safetensors:

from safetensors import safe_open

with safe_open("model.safetensors", framework="pt", device="cpu") as f:
    weight = f.get_tensor("u1.enc.0.weight")

Reproduce the conversion with the original checkpoint and the frcrn_se source package from ClearerVoice-Studio / clearvoice 0.1.2:

python convert_frcrn.py last_best_checkpoint.pt output \
  --source-dir ClearerVoice-Studio/clearvoice/clearvoice/models/frcrn_se \
  --revision 3766e6a64b0d8cb58f08d913d617bf129f11ed53

Conversion needs torch, numpy and safetensors. The script also runs a deterministic CPU comparison against the original model on eight synthetic spectra. --validation-npz accepts an optional private set of normalized spectra; input data is never copied into the output directory.

Checks

The source checkpoint matches the original repository at 3766e6a64b0d8cb58f08d913d617bf129f11ed53, SHA256 b22256adbb91b68cf5a3db8f6657a4fb17066eecd5f069803e59c186c1cf3ebb.

CPU comparisons use float32 operations. The bf16 rows reload the actual serialized weights into float32 modules, isolating weight-conversion error from arithmetic precision. Each set contains eight windows of 51 frames. SNR is 10 log10(sum(reference²) / sum(error²)).

equivalence.json contains the measurements. Synthetic spectra are independent random values and are not an audio-quality benchmark. The held-out data is not distributed. These checks measure agreement with the original CPU model; they do not measure task quality or device speed.

License

Apache-2.0. The original checkpoint card declares Apache-2.0 and has no LICENSE file; LICENSE contains the canonical agreement. NOTICE identifies the original repository, revision and changes. convert_frcrn.py is by Xylolabs under Apache-2.0. Not affiliated with or endorsed by Tenstorrent.