A developer fine-tuned the Parakeet speech-to-text model on 720 recorded Harvard sentences to achieve 6.14% word error rate, improving over Whisper Turbo's performance. After discovering the model failed on real-world audio with silence, they augmented the training dataset by adding silence padding to recordings, reducing test WER to 5.29%.
Oído is an open-vocabulary speech recognition system running on a $5 ESP32-S3 microcontroller with no cloud or neural accelerator, achieving 3.7% word error rate on LibriSpeech—reportedly the most accurate result published for any microcontroller. The system uses a Conformer CTC model optimized with int8 quantization and custom SIMD kernels for the ESP32's PIE vector unit.