Repository navigation
What's new
Models now come from the source. Parakeet regenerates its own ASR models from the official NVIDIA checkpoint, so the server no longer depends on a third-party ONNX conversion.
Local attention for long audio. The encoder now runs with local attention (rel_pos_local_attn). Long audio improves: 1.9% vs 2.9% WER on a 29-minute reference.