Mel Feature
A speech feature representation that mimics the nonlinear frequency perception of the human ear. It uses short-time FFT and triangular Mel filter banks to compress the spectrum into a low-dimensional feature map, and is a common acoustic feature in speech recognition and keyword detection.
Timeline (last 90 days)
Before migrating a model to another runtime, you must verify the sample rate, PCM format, frame length and frame shift, Mel feature configuration, input tensor shape, output label order, quantization settings, and detection thresholds; even if the model file is unchanged, recall can drop as soon as preprocessing changes.
Mel features are obtained by first performing an FFT on short-time audio frames (e.g., 25ms window, 10ms step) to get a spectrum, then compressing it through a triangular Mel filter bank to 20-40 frequency bands, and finally taking the logarithm to obtain the feature map.
Once the Mel feature parameters (sample rate, frame length, number of Mel bands) are determined, training and inference must be strictly consistent, otherwise model accuracy will drop noticeably.
All Facts (3)
Mel features are obtained by first performing an FFT on short-time audio frames (e.g., 25ms window, 10ms step) to get a spectrum, then compressing it through a triangular Mel filter bank to 20-40 frequency bands, and finally taking the logarithm to obtain the feature map.
60%UnverifiedBefore migrating a model to another runtime, you must verify the sample rate, PCM format, frame length and frame shift, Mel feature configuration, input tensor shape, output label order, quantization settings, and detection thresholds; even if the model file is unchanged, recall can drop as soon as preprocessing changes.
50%UnverifiedOnce the Mel feature parameters (sample rate, frame length, number of Mel bands) are determined, training and inference must be strictly consistent, otherwise model accuracy will drop noticeably.
50%