[feat] Add performance and separation quality benchmarks

[dhunstack]

Add benchmarking for PyTorch Demucs and C++ ONNX scripts -

• Separation quality using SI-SDR
• Performance metrics on CPU
• Performance metrics on GPU

CPU Performance Comparison

Processing Speed

Metric	PyTorch Model	C++ ONNX Model	Improvement
Total Processing Time	5,380.35 sec	4,415.30 sec	17.94% faster
Processing Time for 1 min input	25.89 sec	21.24 sec	17.94% faster

Audio Quality (SI-SDR in dB)

Stem	PyTorch Model (dB)	C++ ONNX Model (dB)	Difference
drums	9.50	9.48	-0.02
bass	7.87	7.80	-0.07
other	4.75	4.66	-0.09
vocals	7.90	7.82	-0.08
Overall	7.50	7.44	-0.06

Summary Statistics

Dataset	PyTorch Model	C++ ONNX Model
Total Tracks	50	50
Total Audio Duration	12,470.98 sec (~3.46 hours)	12,470.98 sec (~3.46 hours)
Model Identifier	htdemucs	../onnx-models/htdemucs.ort

Key Findings

• Performance: The C++ ONNX model is 17.94% faster than the PyTorch implementation
• Quality: Audio separation quality is nearly identical, with only minor differences (< 0.1 dB) in SI-SDR scores
• Consistency: Both models processed the same 50 tracks from the dataset with no failures

The results demonstrate that the ONNX export successfully maintains the audio quality while providing performance improvements for deployment scenarios.

GPU Performance Comparison

Unlike the CPU comparison which uses C++, this ONNX inference was done on Python using ONNXRuntime.

Processing Speed

Metric	PyTorch Model (GPU)	Python ONNXRuntime (GPU)	Difference
Total Processing Time	354.13 sec	386.40 sec	8.35% slower
Processing Time for 1 min input	1.70 sec	1.86 sec	8.35% slower

Audio Quality (SI-SDR in dB)

Stem	PyTorch Model (dB)	Python ONNXRuntime (dB)	Difference
drums	9.49	9.50	+0.01
bass	7.77	7.88	+0.11
other	4.72	4.76	+0.04
vocals	7.85	7.87	+0.02
Overall	7.46	7.50	+0.04

Summary Statistics

Dataset	PyTorch Model (GPU)	Python ONNXRuntime (GPU)
Total Tracks	50	50
Total Audio Duration	12,470.98 sec (~3.46 hours)	12,470.98 sec (~3.46 hours)
Model Identifier	htdemucs	htdemucs

Key Findings

• Performance: The PyTorch model is 8.35% faster than the ONNX model run on Python.
• Quality: Audio separation quality is very similar, with the ONNX model showing slightly better SI-SDR scores (+0.04 dB overall)
• Consistency: Both models processed the same 50 tracks from the dataset with no failures

The results show that on GPU, PyTorch maintains a performance advantage while both models deliver comparable audio quality.