soundsense-kwt-tiny

SoundSense AI - KWT-Tiny Keyword Spotter

This is a fine-tuned/adapted model checkpoint for SoundSense AI.

Base model / source:

• Keyword Transformer (KWT), https://arxiv.org/abs/2104.00769
• Custom KWT-Tiny implementation (12-layer transformer, 64-dim embeddings)

Training data:

• Google Speech Commands v2 (35 keyword classes + silence + unknown, 37 total classes)

Use:

• Part of SoundSense AI hackathon submission (Stage 5: Keyword Spotting).
• Runs with shared weights across all 3 separated speaker streams for simultaneous multi-user command word detection, at zero additional parameter cost.

Limitations:

• Built for prototype/demo use.
• Trained on isolated English keyword clips, not continuous Hinglish speech or custom wake phrases (e.g. "hey jarvis"). Best suited for single-word command detection within an already-separated, already-VAD-gated audio segment.
• Performance should be verified on the target environment before deployment.