Notes for Inference

Our current inference pipeline supports single input only and does not support batch processing.

We provide two inference modes: text only and text & speech. You can set the decode_text_only parameter in the inference script to choose your preferred mode.

If using CosyVoice for decoding (as employed in SLAM-Omni), please take note of the following:

• Download the corresponding CosyVoice-300M-SFT model from CosyVoice and set the codec_decoder_path parameter in your script to its location.
• You can customize the output voice tone by specifying the audio_prompt_path. A selection of optional voices is provided in the prompt directory. If not specified, the default voice tone will be used.