ReConSE

Model Summary

ReConSE: Universal speech enhancement via recurrent control over a pretrained F5-TTS model.

Model Weights

Two checkpoint files are provided:

• model_70000_inference.pt — Full inference checkpoint. It contains the EMA weights for the complete ReConSE model: the F5-TTS v1 Base backbone, Recurrent ControlNet, and Qwen audio adapters. It excludes optimizer and scheduler states and the non-EMA training weights. The Qwen audio encoder and Vocos vocoder are not included.

• model_70000_controlnet_adapter.pt — Compact EMA weights for the Recurrent ControlNet and Qwen audio adapters only. It does not contain the F5-TTS backbone, so it is not a standalone checkpoint.

Required External Assets

Both checkpoint variants require the ReConSE code and configuration from the GitHub repository, including the model configuration and vocabulary file at src/f5_tts/infer/examples/vocab.txt.

Because both checkpoints use Qwen audio adapters, download the Qwen3-ASR-1.7B model for its audio encoder. The Qwen encoder weights are not bundled in either ReConSE checkpoint.

The Vocos vocoder is also required to convert the generated mel spectrograms into audio. Download Vocos mel-24kHz.

Using model_70000_inference.pt

This checkpoint already contains the F5-TTS backbone weights. Load it with the ReConSE inference code and provide the external Qwen encoder and Vocos assets through the configuration.

Using model_70000_controlnet_adapter.pt

This checkpoint requires the F5-TTS v1 Base model weights, including model_1250000.safetensors.

The current inference loader constructs the model architecture and then loads the supplied checkpoint. Since this file contains only ControlNet and adapter weights, using it directly would leave the F5-TTS backbone uninitialized. Modify the loading code to:

1. Initialize the F5-TTS v1 Base backbone from its pretrained checkpoint.
2. Load the ReConSE ControlNet and adapter weights on top of that backbone.

The vocabulary used to construct the model must match the vocabulary expected by the checkpoint.

Code

GitHub repository