Spaces:
Running on Zero
Running on Zero
File size: 2,047 Bytes
875e4af | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 | {
"model_name": "whisper-tiny-frozen-encoder-logreg",
"architecture": [
"16kHz mono audio",
"Whisper Tiny frozen encoder (openai/whisper-tiny, loaded via transformers at runtime, NOT bundled here)",
"mean pooling over encoder time axis (384-dim vector for whisper-tiny)",
"Logistic Regression (StandardScaler -> LogisticRegression, class_weight=balanced, C=1.0)",
"threshold/hysteresis decision layer (src/turn_detector/inference.py: TurnDecisionConfig)"
],
"classifier_artifact": "whisper_classifier.joblib",
"classifier_trained_on": {
"experiment": "EXP-004",
"run_location": "Google Colab (notebooks/EXP004_whisper_baseline.ipynb)",
"sample": "Fresh 250-clip HF-streaming sample from pipecat-ai/smart-turn-data-v3.2-train, independent of the Phase 3 sample",
"dev_n": 175,
"val_n": 75,
"seed": 42
},
"measured_performance_val_set": {
"n": 75,
"f1": 0.693,
"accuracy": 0.693,
"false_end_rate": 0.257,
"false_continue_rate": 0.350,
"note": "Measured on ONE 75-clip validation split from ONE 250-clip sample. Not full-dataset or official-test-set performance. See docs/RESULTS.md."
},
"encoder_param_count": 8208384,
"encoder_approx_size_fp32_mb": 32.8,
"classifier_param_count": 92,
"classifier_approx_size_bytes": 2920,
"sklearn_version_trained_with": "1.6.1",
"sklearn_version_note": "This classifier was pickled with scikit-learn 1.6.1 in the Colab environment. Loading it with a different sklearn version (e.g. this project's dev sandbox, which has 1.8.0) raises an InconsistentVersionWarning — observed and non-fatal in testing, but worth pinning scikit-learn==1.6.1 in a production deployment to avoid any risk of silent incompatibility.",
"whisper_weights_not_included": true,
"whisper_weights_loading": "TurnDetector loads openai/whisper-tiny via transformers.WhisperModel.from_pretrained() at runtime, which downloads/caches the weights (~151MB, separate from this classifier). Requires network access to Hugging Face on first run."
}
|