test_outputs/README_tests.md

Supertonic Quantized – Test Outputs

This folder was generated automatically by a Colab script that:

1. Downloaded the Hugging Face repo Shadow0482/supertonic-quantized
2. Located all *.onnx models (both fp16/ and int8_dynamic/)
3. Ran each model once with dummy inputs using ONNX Runtime
4. Saved:
   • .wav files for audio-like tensors (1D or 2D, 1–2 channels, >=16 samples)
   • .npy files for all other outputs

All paths below are relative to the test_outputs/ directory.

Per-model results

• fp16/duration_predictor.fp16.onnx -> fp16/fp16_duration_predictor.fp16_out0.npy
• fp16/text_encoder.fp16.onnx -> fp16/fp16_text_encoder.fp16_out0.npy
• fp16/vector_estimator.fp16.onnx -> FAILED
• fp16/vocoder.fp16.onnx -> fp16/fp16_vocoder.fp16_out0.wav
• int8_dynamic/duration_predictor.int8.onnx -> int8_dynamic/int8_dynamic_duration_predictor.int8_out0.npy
• int8_dynamic/text_encoder.int8.onnx -> int8_dynamic/int8_dynamic_text_encoder.int8_out0.npy
• int8_dynamic/vector_estimator.int8.onnx -> int8_dynamic/int8_dynamic_vector_estimator.int8_out0.npy
• int8_dynamic/vocoder.int8.onnx -> int8_dynamic/int8_dynamic_vocoder.int8_out0.wav

Models that failed to load / run

• fp16/vector_estimator.fp16.onnx -> FAILED: [ONNXRuntimeError] : 1 : FAIL : Load model from /content/supertonic/supertonic_quantized/fp16/vector_estimator.fp16.onnx failed:Type Error: Type (tensor(float16)) of output arg (/vector_field/main_blocks.3/attn/Cast_output_0) of node (/vector_field/main_blocks.3/attn/Cast) does not match expected type (tensor(float)).

Note: These tests use synthetic dummy inputs. They confirm that the quantized ONNX graphs load and execute, but they are not a replacement for real end-to-end TTS quality evaluation.