Hi Unity team and community
I’m able to load the FireRedASR ONNX model using Sentis, but I get an inference error. However, I can infer using the onnx runtime in Linux.
I’m using the FireRedASR model and exporting it to ONNX using the method or using other already exported ONNX models in Unity.
The following is the error I get when exporting the ONNX model using the method
The input to this model is the raw audio sampling information.
The following is an error message when using another already exported ONNX model
The inputs to this model are x (the audio FBank features) and x_len (the length of the audio). I used the Whisper-Tiny project in Sentis to generate the audio FBank features.（ logmel_spectrogram.onnx）
This model is a speech recognition model based on the Attention Encoder-Decoder.
The Encoder is based on the Conformer.
The Decoder is based on the Transformer-Decoder, using CrossAttention to focus on the Encoder’s output.
The following is an inference result using the same ONNX model, using ONNX runtime==1.22.0 and ONNX==1.18.0. The inference code is based on this.
I believe the Onnx model should be fine.
I’m confused as to why the internal dimensions of this Onnx model don’t match during inference when used in Unity Sentis.
The following link may be helpful 