You cannot find a non-streaming model anywhere accepting audio files of any length.

Why not, you say

Using Fine-tuned Whisper

Official whisper models only accept 30-second audios. To improve the throughput, you could fine-tune the whisper model to remove the 30 seconds restriction. Seeexamples.

We prepared twoChinese fine-tuned whisper TensorRT-LLM weights repo. They could be directly used fromhere.

Theoretically, if your RAM is infinite, then it is possible.

Removing the 30s constraint means you can input audio files less than 30s without padding. It does not mean you.can input files of arbitrary length.

By the way, is there any disadvantage of using a vad model here?

Теоретически, если у вас оперативная память бесконечна, то это возможно.

Снятие ограничения на 30 секунд означает, что вы можете вводить аудиофайлы менее 30 секунд без заполнения. Это не означает, что вы можете вводить файлы произвольной длины.

Кстати, есть ли здесь какие-то недостатки использования модели vad?

Okay, I realized that long files aren't possible.
I haven't noticed any flaws in the VAD model, I didn’t use VAD enough to properly assess the audio.

How do i enter audio files less than 30 seconds without filling in ?

Padding is invisible for users if the input audio is less than 30s. it is an implementation detail.

Padding is invisible for users if the input audio is less than 30s. it is an implementation detail.

Faytuning the model has improved your results ?

strings x = (x + self.positional_embedding[:x.shape[1], :]).to(x.dtype)