NamiGen Language Detection v1
A five-class ONNX language detection model. The class order is Vietnamese,
Japanese, Korean, English, Chinese (vi, ja, ko, en, zh). It is
intended for mono 8 kHz audio with the matching NamiGen service frontend.
The graph receives precomputed normalized log-mel features of shape
[batch, 1, 80, 398] and an int64 valid_samples vector. Supply the genuine
sample count separately when right-padding audio to four seconds. The graph
contains the corrected population BatchNorm buffers. A generic audio pipeline
will not reproduce this frontend automatically; use the accompanying service
implementation and normalization file.
namigen-lang-detection-v1.sha256 records the expected SHA-256 digests of the
graph and normalization data. Verify them before use. Model probabilities and
application thresholds require validation on the intended audio domain; this
package makes no accuracy guarantee.