rado-tech/uzbek-gender-cnn

Uzbek Speaker Gender Recognition (CNN)

A compact convolutional network that predicts speaker gender from Uzbek speech. 60,706 parameters, 0.2 MB on disk, ~4 ms per 2-second window on a laptop CPU.

Model details

• Architecture: four Conv2d → BatchNorm2d → ReLU → MaxPool2d blocks (1→16→32→64→64 channels), global average pooling, dropout 0.3, linear head
• Input: log-mel spectrogram, 64 mel bands × 200 frames (2 s at 16 kHz), normalised with per-band statistics computed on the training split
• Output: two logits; index 0 = Ayol (female), 1 = Erkak (male)
• Training: AdamW, lr 3e-3, cosine schedule, weight decay 1e-4, 15 epochs, batch size 64, random 2-second crops as augmentation
• License: Apache 2.0

Intended use

Exploratory and research use on Uzbek speech: dataset analysis, demos, and as a lightweight baseline for voice-attribute classification.

Out of scope. Do not use this model to make decisions about individuals — access control, identity verification, profiling, content moderation, or any setting where a person is affected by the output. It predicts an acoustic correlate of vocal tract physiology, not a person's gender identity.

Evaluation

Held-out test split: 3,049 clips from 364 speakers disjoint from training and validation.

Metric	Value
Accuracy	98.10%
Precision (male)	0.9940
Recall (male)	0.9681
Recall (female)	0.9940
F1	0.9809

A fine-tuned wav2vec2-base baseline reaches 98.75% on the same split at 126× the latency and 1480× the size.

About 30 of the 58 test errors trace to two contributor accounts with demonstrably wrong labels — one contains recordings from two different people. Excluding them the model scores 99.07%, though the reportable figure remains 98.10%.

Limitations and biases

• Language. Trained only on Uzbek. Accuracy elsewhere is unverified, though the acoustic cues (fundamental frequency, formants) are largely language-independent.
• Binary labels. The source dataset labels gender as binary, so the model does too. This reflects the data, not the range of human gender identity. Voices that do not fit either training class will be misclassified or land in the uncertain band.
• Age distribution. About 75% of training speakers are 18–24. Accuracy was measured across age brackets with no significant effect, but coverage of older and younger voices is thin.
• Regional variation. Namangan-accented speech is the weakest region (96.6%) while the transformer baseline handles it perfectly, suggesting the small model is more accent-sensitive.
• Recording conditions. Trained on crowdsourced phone and laptop recordings. Studio audio or heavy background noise are out of distribution.
• VAD. The bundled voice-activity detector is energy-based and treats loud noise as speech.

Usage

from genderid import GenderClassifier

clf = GenderClassifier("cnn_best.pt")
result = clf.predict_file("sample.wav")

print(result.label)        # 'Ayol' | 'Erkak' | 'uncertain' | 'no_speech'
print(result.confidence)   # None when uncertain
print(result.probability)  # p(male), averaged over speech windows

Long recordings are split into overlapping 2-second windows; silent windows are dropped and the remaining probabilities averaged. When the average falls between 0.35 and 0.65 the model reports uncertain instead of guessing.

Accepts any format ffmpeg can read. For in-memory audio use predict_waveform(array, sample_rate).

Training data

DavronSherbaev/uzbekvoice-filtered, Apache 2.0. A 30,000-clip subset (34.8 hours) from 3,505 speakers, filtered on community quality signals, capped at 40 clips per speaker, and split so that no speaker appears in more than one partition.

Citation

@software{uzbek_gender_cnn,
  title  = {Uzbek Speaker Gender Recognition},
  year   = {2026},
  note   = {CNN trained on uzbekvoice-filtered},
  license = {Apache-2.0}
}