File size: 5,139 Bytes
6b0fab3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
from __future__ import annotations

import argparse
import json
from pathlib import Path

import numpy as np
import onnxruntime as ort
import soundfile as sf
from scipy.signal import resample_poly


MODEL_PATH = Path(__file__).with_name("model.onnx")
SAMPLE_RATE = 16_000
CHUNK_FRAMES = 160_000
SUPPORTED_EXTENSIONS = {".wav", ".flac", ".mp3", ".ogg", ".m4a"}


def load_audio(path: Path) -> tuple[np.ndarray, int]:
    audio, sample_rate = sf.read(path, dtype="float32", always_2d=True)
    waveform = np.ascontiguousarray(audio.mean(axis=1), dtype=np.float32)
    return waveform, int(sample_rate)


def resample_audio(
    waveform: np.ndarray, source_rate: int, target_rate: int
) -> np.ndarray:
    gcd = np.gcd(source_rate, target_rate)
    waveform = resample_poly(
        waveform, target_rate // gcd, source_rate // gcd
    ).astype(np.float32)
    return np.ascontiguousarray(waveform)


def layer_norm(waveform: np.ndarray, eps: float = 1e-5) -> np.ndarray:
    mean = waveform.mean(dtype=np.float64)
    variance = waveform.var(dtype=np.float64)
    return ((waveform - mean) / np.sqrt(variance + eps)).astype(np.float32)


def chunk_waveform(waveform: np.ndarray, chunk_frames: int) -> np.ndarray:
    if chunk_frames <= 0 or waveform.size <= chunk_frames:
        return waveform[None, :]

    chunks = [
        waveform[start : start + chunk_frames]
        for start in range(0, waveform.size, chunk_frames)
    ]
    max_length = max(chunk.size for chunk in chunks)
    batch = np.zeros((len(chunks), max_length), dtype=np.float32)

    for index, chunk in enumerate(chunks):
        batch[index, : chunk.size] = chunk

    return batch


def softmax(logits: np.ndarray) -> np.ndarray:
    logits = logits.astype(np.float64)
    probabilities = np.exp(logits - logits.max())
    return probabilities / probabilities.sum()


class TTSSuitabilityClassifier:
    def __init__(
        self,
        model_path: str | Path = MODEL_PATH,
        provider: str = "auto",
        cuda_device_id: int = 0,
    ) -> None:
        available = set(ort.get_available_providers())

        if provider == "auto":
            provider = "cuda" if "CUDAExecutionProvider" in available else "cpu"

        if provider == "cuda":
            if "CUDAExecutionProvider" not in available:
                raise RuntimeError(
                    "CUDAExecutionProvider is unavailable. Install onnxruntime-gpu "
                    "or use provider='cpu'."
                )
            providers = [
                ("CUDAExecutionProvider", {"device_id": cuda_device_id}),
                "CPUExecutionProvider",
            ]
        elif provider == "cpu":
            providers = ["CPUExecutionProvider"]
        else:
            raise ValueError("provider must be one of: auto, cpu, cuda")

        self.session = ort.InferenceSession(str(model_path), providers=providers)
        self.input_name = self.session.get_inputs()[0].name
        self.output_names = [output.name for output in self.session.get_outputs()]

    def predict(self, audio_path: str | Path) -> dict[str, object]:
        path = Path(audio_path).expanduser().resolve()
        waveform, sample_rate = load_audio(path)

        if sample_rate != SAMPLE_RATE:
            waveform = resample_audio(waveform, sample_rate, SAMPLE_RATE)

        waveform = layer_norm(waveform)
        batch = chunk_waveform(waveform, CHUNK_FRAMES)
        logits = self.session.run(
            self.output_names, {self.input_name: batch}
        )[0].mean(axis=0)
        probabilities = softmax(logits)
        predicted_class = int(probabilities.argmax())

        return {
            "path": str(path),
            "label": "tts" if predicted_class == 1 else "not_tts",
            "predicted_class": predicted_class,
            "p_not_tts": float(probabilities[0]),
            "p_tts": float(probabilities[1]),
            "logits": [float(value) for value in logits],
        }


def collect_audio_paths(path: Path) -> list[Path]:
    path = path.expanduser().resolve()
    if path.is_file():
        return [path]

    return sorted(
        child
        for child in path.rglob("*")
        if child.is_file() and child.suffix.lower() in SUPPORTED_EXTENSIONS
    )


def main() -> None:
    parser = argparse.ArgumentParser(
        description="ONNX inference for the TTS suitability classifier."
    )
    parser.add_argument("audio", type=Path, help="Audio file or directory.")
    parser.add_argument(
        "--model", type=Path, default=MODEL_PATH, help="Path to model.onnx."
    )
    parser.add_argument(
        "--provider", choices=("auto", "cpu", "cuda"), default="auto"
    )
    parser.add_argument("--cuda-device-id", type=int, default=0)
    args = parser.parse_args()

    classifier = TTSSuitabilityClassifier(
        args.model, args.provider, args.cuda_device_id
    )
    paths = collect_audio_paths(args.audio)

    if not paths:
        raise RuntimeError(f"No supported audio files found at '{args.audio}'.")

    for path in paths:
        print(json.dumps(classifier.predict(path), ensure_ascii=False))


if __name__ == "__main__":
    main()