TF-Restormer baseline

This repository contains the corrected public TF-Restormer baseline checkpoint. It performs universal speech restoration and bandwidth extension up to 48 kHz.

Corrected checkpoint

The model was replaced on 2026-08-09 after GitHub issue #2 showed that the previous public artifact emitted almost no signal above 8 kHz.

  • Source training checkpoint: epoch.0019.pth
  • Exported model SHA-256: d67115273bed06cb1ab06de57168e6ce6e56dfcf3c626b3e3fbc6eb945b5424e
  • Raw source checkpoint SHA-256: f41a075ee3edf4c747f7626a0c64b09c361cd8a3058201d24f6572313c3501a4
  • Config SHA-256: d94e7509dfa84a09d76a81a71a4861fc8cf3684b753d45eb24b7ff5ef27daf8e

The exported state has 323 tensors and is bit-exact with epoch 19 after removal of training-only profiling buffers. The bundled config enables input-bandwidth detection.

What happened to the previous artifact?

The old Hub file was an automatic export of the latest checkpoint, epoch 25. Epochs 20โ€“25 had been resumed under changed data/config conditions involving mixed-sampling-rate, band-limited clean targets. The estimator consequently learned to suppress high-frequency output. The collapse starts at epoch 20 and grows monotonically through epoch 25.

Two sample-rate handling bugs in auxiliary losses were also fixed in the source repository, but controlled evaluation showed that they did not cause this checkpoint collapse.

Usage

from tf_restormer import SEInference

model = SEInference.from_pretrained(
    checkpoint_path="shinuh/tf-restormer-baseline",
    device="cuda",
)

result = model.process_file(
    "noisy_16k.wav",
    output_path="restored_48k.wav",
    fs_out=48000,
)

Pass input audio at its native sampling rate. Do not pre-upsample 8/16 kHz content before inference; the model performs bandwidth extension internally.

Release verification

Three fixed 16โ†’48 kHz demo samples passed state-integrity, config, and inference checks. Their mean 8โ€“20 kHz / 0.3โ€“3.4 kHz energy ratio was โˆ’12.90 dB; the same test rejects the collapsed public epoch-25 behavior. The uploaded files are downloaded and verified again as part of the release procedure.

Source and examples: https://github.com/dmlguq456/TF_Restormer

Downloads last month
7
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support