File size: 8,253 Bytes
2d70679
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
"""Dependency-light audio normalization and log-mel feature extraction.

The frontend configuration is serialized next to every exported model. Keeping
one implementation for export validation and inference prevents a very common
failure mode: a correct ONNX graph fed subtly different features in production.
"""

from __future__ import annotations

from dataclasses import asdict, dataclass
from functools import lru_cache
from typing import Any


def _numpy() -> Any:
    try:
        import numpy as np
    except ImportError as exc:  # pragma: no cover - exercised in minimal installs
        raise RuntimeError("Audio inference requires numpy; install the base package") from exc
    return np


@dataclass(frozen=True, slots=True)
class FrontendConfig:
    sample_rate: int = 16_000
    max_seconds: float = 8.0
    n_fft: int = 400
    win_length: int = 400
    hop_length: int = 160
    n_mels: int = 80
    f_min: float = 0.0
    f_max: float = 8_000.0
    normalization: str = "whisper"
    pad_side: str = "left"

    def __post_init__(self) -> None:
        if self.sample_rate <= 0 or self.max_seconds <= 0:
            raise ValueError("sample_rate and max_seconds must be positive")
        if self.n_fft <= 0 or self.win_length <= 0 or self.hop_length <= 0:
            raise ValueError("FFT and window sizes must be positive")
        if self.win_length > self.n_fft:
            raise ValueError("win_length cannot exceed n_fft")
        if self.n_mels <= 0:
            raise ValueError("n_mels must be positive")
        if not 0.0 <= self.f_min < self.f_max <= self.sample_rate / 2:
            raise ValueError("mel frequency bounds must lie inside Nyquist")
        if self.normalization not in {"whisper", "log10", "none"}:
            raise ValueError("unsupported normalization")
        if self.pad_side not in {"left", "right"}:
            raise ValueError("pad_side must be 'left' or 'right'")

    @property
    def max_samples(self) -> int:
        return round(self.sample_rate * self.max_seconds)

    @property
    def target_frames(self) -> int:
        return self.max_samples // self.hop_length

    def to_dict(self) -> dict[str, Any]:
        return asdict(self)


def normalize_waveform(audio: Any) -> Any:
    """Convert mono/stereo integer/float audio to finite mono float32 in [-1, 1]."""

    np = _numpy()
    samples = np.asarray(audio)
    if samples.size == 0:
        raise ValueError("audio cannot be empty")
    original_dtype = samples.dtype
    if samples.ndim == 2:
        # Gradio commonly returns [samples, channels]; accept [channels, samples] too.
        channel_axis = 1 if samples.shape[1] <= 8 else 0
        samples = samples.astype(np.float32).mean(axis=channel_axis)
    elif samples.ndim != 1:
        raise ValueError(f"expected mono/stereo audio, got shape {samples.shape}")
    if np.issubdtype(original_dtype, np.integer):
        info = np.iinfo(original_dtype)
        scale = float(max(abs(info.min), info.max))
        samples = samples.astype(np.float32) / scale
    else:
        samples = samples.astype(np.float32, copy=False)
    samples = np.nan_to_num(samples, nan=0.0, posinf=1.0, neginf=-1.0)
    peak = float(np.max(np.abs(samples)))
    if peak > 1.0:
        samples = samples / peak
    return np.clip(samples, -1.0, 1.0)


def resample_waveform(audio: Any, source_rate: int, target_rate: int) -> Any:
    """Resample with the same deterministic linear rule used during training.

    Optional packages must not change model inputs, so this deliberately avoids
    a SciPy-dependent branch.  Higher-quality telephony resampling can happen
    upstream, but training and serving always agree at this boundary.
    """

    np = _numpy()
    if source_rate <= 0 or target_rate <= 0:
        raise ValueError("sample rates must be positive")
    samples = normalize_waveform(audio)
    if source_rate == target_rate:
        return samples
    output_length = max(1, round(len(samples) * target_rate / source_rate))
    old_x = np.linspace(0.0, 1.0, len(samples), endpoint=False)
    new_x = np.linspace(0.0, 1.0, output_length, endpoint=False)
    return np.interp(new_x, old_x, samples).astype(np.float32)


def pad_or_trim(audio: Any, config: FrontendConfig) -> tuple[Any, int]:
    """Return fixed-length audio and number of genuine (non-padding) samples."""

    np = _numpy()
    samples = normalize_waveform(audio)
    if len(samples) >= config.max_samples:
        return samples[-config.max_samples :].copy(), config.max_samples
    pad = config.max_samples - len(samples)
    widths = (pad, 0) if config.pad_side == "left" else (0, pad)
    return np.pad(samples, widths).astype(np.float32), len(samples)


def _hz_to_mel(value: Any) -> Any:
    np = _numpy()
    return 2595.0 * np.log10(1.0 + np.asarray(value) / 700.0)


def _mel_to_hz(value: Any) -> Any:
    np = _numpy()
    return 700.0 * (10.0 ** (np.asarray(value) / 2595.0) - 1.0)


@lru_cache(maxsize=16)
def mel_filterbank(config: FrontendConfig) -> Any:
    """Create a deterministic triangular mel filter bank."""

    np = _numpy()
    mel_points = np.linspace(_hz_to_mel(config.f_min), _hz_to_mel(config.f_max), config.n_mels + 2)
    hz_points = _mel_to_hz(mel_points)
    fft_hz = np.linspace(0.0, config.sample_rate / 2, config.n_fft // 2 + 1)
    filters = np.zeros((config.n_mels, len(fft_hz)), dtype=np.float32)
    for index in range(config.n_mels):
        left, center, right = hz_points[index : index + 3]
        filters[index] = np.maximum(
            0.0,
            np.minimum(
                (fft_hz - left) / max(center - left, 1e-12),
                (right - fft_hz) / max(right - center, 1e-12),
            ),
        )
    # Area normalization reduces frequency-dependent scale drift.
    enorm = 2.0 / np.maximum(hz_points[2 : config.n_mels + 2] - hz_points[: config.n_mels], 1e-12)
    result = filters * enorm[:, None]
    result.flags.writeable = False
    return result


@lru_cache(maxsize=16)
def _hann_window(length: int) -> Any:
    np = _numpy()
    window = np.hanning(length).astype(np.float32)
    window.flags.writeable = False
    return window


def log_mel_spectrogram(
    audio: Any,
    source_rate: int,
    config: FrontendConfig | None = None,
) -> tuple[Any, Any]:
    """Return ``[n_mels, frames]`` features and a valid-frame mask.

    The implementation follows Whisper's centered-STFT and dynamic-range
    normalization convention closely, while remaining free of torch/librosa at
    inference time. Export parity tests compare it against the training path.
    """

    np = _numpy()
    cfg = config or FrontendConfig()
    if source_rate <= 0:
        raise ValueError("source sample rate must be positive")
    # Bound work before resampling: an uploaded meeting can be hours long, while
    # endpoint intent uses only the configured suffix. Training uses the same rule.
    normalized = normalize_waveform(audio)
    source_suffix_samples = max(1, round(cfg.max_seconds * source_rate))
    normalized = normalized[-source_suffix_samples:]
    resampled = resample_waveform(normalized, source_rate, cfg.sample_rate)
    fixed, valid_samples = pad_or_trim(resampled, cfg)

    pad = cfg.n_fft // 2
    padded = np.pad(fixed, (pad, pad), mode="reflect")
    frames = np.lib.stride_tricks.sliding_window_view(padded, cfg.win_length)[:: cfg.hop_length]
    frames = frames[: cfg.target_frames]
    window = _hann_window(cfg.win_length)
    spectrum = np.fft.rfft(frames * window[None, :], n=cfg.n_fft, axis=1)
    power = (spectrum.real**2 + spectrum.imag**2).astype(np.float32)
    mel = np.maximum(mel_filterbank(cfg) @ power.T, 1e-10)
    features = np.log10(mel)
    if cfg.normalization == "whisper":
        features = np.maximum(features, float(features.max()) - 8.0)
        features = (features + 4.0) / 4.0
    elif cfg.normalization == "none":
        features = mel

    frame_mask = np.zeros(cfg.target_frames, dtype=np.float32)
    valid_frames = min(
        cfg.target_frames, max(1, (valid_samples + cfg.hop_length - 1) // cfg.hop_length)
    )
    if cfg.pad_side == "left":
        frame_mask[-valid_frames:] = 1.0
    else:
        frame_mask[:valid_frames] = 1.0
    return features.astype(np.float32), frame_mask