Upgrade to Silero-Vad V5

[hoonlight]

https://github.com/snakers4/silero-vad/releases/tag/v5.0

• The V5 model now only works with a fixed size window, so the window_size_samples parameter is removed and its value is fixed at 512.
• Change to use the state variable instead of the existing h and c variables.
• Slightly changed internal logic, now some context (part of previous chunk) is passed along with the current chunk.
• Change the dimensions of the state variable from 64 to 128.
• Replace ONNX file with V5 version

[trungkienbkhn]

@hoonlight , thanks for quickly adapting Silero-Vad V5 for fw after this model was released. Have you run benchmarks for it yet?

[hoonlight]

@hoonlight , thanks for quickly adapting Silero-Vad V5 for fw after this model was released. Have you run benchmarks for it yet?

No, I haven't run the benchmarks yet.
I found it a little while ago, thanks.

However, since I won't have access to a GPU for a while, I should be able to run the benchmarks in a couple weeks.

[Purfview]

Thanks for adapting v5, I was just thinking to do it, good that I noticed your PR. 😄

Btw, about GPU, look there #499 (comment)

[trungkienbkhn]

For information, I ran benchmarks with GPU H100 and large-v3 model.
Below are the results:

1. Speed benchmark:
Processing audio with duration 13:19.231s
Detected language 'fr' with probability 1.00

System	Min execution time
Faster-Whisper	41.413s
FW with SILERO VAD V5	39.529s

2. WER benchmark:
Dataset: librispeech_asr
Number of audio used for evaluation: 500

System	WER
Faster-Whisper	3.139
FW with SILERO VAD V5	2.815

3. Memory benchmark:
GPU name: NVIDIA H100 PCIe
GPU device index: 0

System	Maximum increase of RAM	Maximum GPU memory usage	Maximum GPU power usage
Faster-Whisper	1222 MiB	5107MiB / 81559MiB	145W / 350W
FW with SILERO VAD V5	1225 MiB	5107MiB / 81559MiB	149W / 350W

The results look good to me. Speed has improved a bit, as described in the VAD V5 model release:

3x faster inference for TorchScript, 10% faster inference for ONNX;