snakers4/silero-vad

AI Summary: This GitHub issue seeks guidance on data preparation for fine-tuning a model. The user specifically asks about the optimal amount of data (in hours) required for effective fine-tuning and whether public datasets are necessary. If public data is needed, the user also inquires about the recommended ratio of public to private data during the training process.

AI Summary: The user is building an Android wake word detection pipeline where Silero VAD acts as a pre-filter for an `openWakeWord` model. The challenge arises because VAD processes small 520-sample audio chunks, but the wake word model strictly requires larger 1280-sample inputs. The user is asking if Silero VAD v6 can be configured to process these larger chunks directly to resolve the input size mismatch.

AI Summary: This GitHub issue is a user's question seeking clarification on whether the Voice Activity Detection (VAD) system supports filtering voices based on proximity. Specifically, the user wants to know if the VAD can detect only nearby voices while ignoring background conversations from people further away.

AI Summary: The user is encountering a CUDA out of memory error specifically during the validation phase while fine-tuning a model on an Nvidia V100, despite the training phase proceeding normally with the default `config.yaml`. They are seeking assistance to understand the potential reasons for this issue.

AI Summary: A user is asking for clarification on the intended use case for the `silero_vad_16k_op15.onnx` model, specifically whether it's preferable for 16kHz audio inputs and systems compatible with ONNX Opset 15, compared to the standard `silero_vad.onnx` model.

AI Summary: The task is to debug a RuntimeError in a PyTorch training loop for a Voice Activity Detector (VAD) model. The error, 'RuntimeError: Trying to backward through the graph a second time', suggests a missing `optimizer.zero_grad()` call before `loss.backward()`. The solution involves adding this call to reset the gradients before each backward pass to prevent accumulation of gradients across iterations.