File size: 10,921 Bytes
adde09b | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 | """High-level Speaker Encoder API.
Provides a unified interface for extracting speaker embeddings from audio
files using either a PyTorch checkpoint or an ONNX model.
Usage (ONNX -- recommended for inference):
encoder = SpeakerEncoder.from_onnx("speaker_encoder.onnx")
emb = encoder.encode("audio.wav")
Usage (PyTorch):
encoder = SpeakerEncoder.from_pytorch("speaker_encoder.ckpt")
emb = encoder.encode("audio.wav")
"""
from __future__ import annotations
import logging
from pathlib import Path
from typing import TYPE_CHECKING
import numpy as np
from .audio_utils import (
DEFAULT_FMAX,
DEFAULT_FMIN,
DEFAULT_HOP_LENGTH,
DEFAULT_N_FFT,
DEFAULT_N_MELS,
DEFAULT_SR,
compute_mel_spectrogram,
load_audio,
normalize_audio,
)
if TYPE_CHECKING:
import onnxruntime
import torch
_LOGGER = logging.getLogger(__name__)
def _infer_hparams(state_dict: dict) -> dict:
"""Infer ECAPA-TDNN hyperparameters from a state_dict.
Examines key tensor shapes to determine input_dim, channels, emb_dim,
and scale so that the model can be reconstructed without explicit
configuration.
Args:
state_dict: Model state dictionary.
Returns:
Dict of keyword arguments for :class:`ECAPATDNN`.
"""
# layer1.0.weight has shape (channels, input_dim, kernel_size)
layer1_weight = state_dict.get("layer1.0.weight")
if layer1_weight is None:
raise ValueError(
"Cannot infer hparams: 'layer1.0.weight' not found in state_dict"
)
channels = layer1_weight.shape[0]
input_dim = layer1_weight.shape[1]
# fc.weight has shape (emb_dim, channels * 2)
fc_weight = state_dict.get("fc.weight")
if fc_weight is None:
raise ValueError("Cannot infer hparams: 'fc.weight' not found in state_dict")
emb_dim = fc_weight.shape[0]
# Infer Res2Net scale from the number of conv modules in layer2.res2net.convs
# convs has (scale - 1) entries: convs.0, convs.1, ..., convs.(scale-2)
scale_minus_1 = 0
for key in state_dict:
if key.startswith("layer2.res2net.convs.") and key.endswith(".weight"):
scale_minus_1 += 1
scale = scale_minus_1 + 1 if scale_minus_1 > 0 else 8
# Infer SE bottleneck from layer2.se.se.1.weight shape (bottleneck, channels, 1)
se_weight = state_dict.get("layer2.se.se.1.weight")
se_bottleneck = se_weight.shape[0] if se_weight is not None else 128
return {
"input_dim": input_dim,
"channels": channels,
"emb_dim": emb_dim,
"scale": scale,
"se_bottleneck": se_bottleneck,
}
class SpeakerEncoder:
"""Speaker Encoder high-level API.
Loads a PyTorch or ONNX speaker encoder model and provides methods
to extract 256-dimensional speaker embeddings from audio files.
Do not instantiate directly; use :meth:`from_pytorch` or :meth:`from_onnx`.
"""
def __init__(self) -> None:
self._mode: str = "none" # "pytorch" or "onnx"
self._pytorch_model: torch.nn.Module | None = None
self._pytorch_device: str = "cpu"
self._onnx_session: onnxruntime.InferenceSession | None = None
@classmethod
def from_pytorch(
cls,
checkpoint_path: str | Path,
device: str = "cpu",
) -> SpeakerEncoder:
"""Load a speaker encoder from a PyTorch checkpoint.
The checkpoint should contain either:
- A raw state_dict (keys like ``layer1.0.weight``), or
- A dict with a ``"model_state_dict"`` key.
Args:
checkpoint_path: Path to the ``.ckpt`` or ``.pt`` file.
device: Torch device string (default: ``"cpu"``).
Returns:
Configured :class:`SpeakerEncoder` instance.
"""
import torch # noqa: PLC0415
from .ecapa_tdnn import ECAPATDNN # noqa: PLC0415
checkpoint_path = Path(checkpoint_path)
if not checkpoint_path.exists():
raise FileNotFoundError(f"Checkpoint not found: {checkpoint_path}")
ckpt = torch.load(str(checkpoint_path), map_location=device, weights_only=True)
if isinstance(ckpt, dict) and "model_state_dict" in ckpt:
state_dict = ckpt["model_state_dict"]
elif isinstance(ckpt, dict) and all(isinstance(k, str) for k in ckpt.keys()):
state_dict = ckpt
else:
raise ValueError(
"Checkpoint format not recognised. Expected a state_dict or a "
"dict with 'model_state_dict' key."
)
# Infer model hyperparameters from the state_dict shapes
hparams = _infer_hparams(state_dict)
model = ECAPATDNN(**hparams)
model.load_state_dict(state_dict)
model.eval()
model.to(device)
encoder = cls()
encoder._mode = "pytorch"
encoder._pytorch_model = model
encoder._pytorch_device = device
_LOGGER.info(
"Loaded PyTorch speaker encoder from %s (device=%s)",
checkpoint_path,
device,
)
return encoder
@classmethod
def from_onnx(cls, onnx_path: str | Path) -> SpeakerEncoder:
"""Load a speaker encoder from an ONNX model.
Uses the project's shared ORT session utilities for optimised
session creation with caching support.
Args:
onnx_path: Path to the ``.onnx`` file.
Returns:
Configured :class:`SpeakerEncoder` instance.
"""
onnx_path = Path(onnx_path)
if not onnx_path.exists():
raise FileNotFoundError(f"ONNX model not found: {onnx_path}")
from ..ort_utils import create_session_with_cache # noqa: PLC0415
session = create_session_with_cache(onnx_path, device="cpu")
encoder = cls()
encoder._mode = "onnx"
encoder._onnx_session = session
_LOGGER.info("Loaded ONNX speaker encoder from %s", onnx_path)
return encoder
# ------------------------------------------------------------------
# Public API
# ------------------------------------------------------------------
def encode(self, audio_path: str | Path) -> np.ndarray:
"""Extract a 256-dimensional speaker embedding from an audio file.
The audio is loaded, peak-normalized, converted to a log-mel
spectrogram, and passed through the encoder model.
Args:
audio_path: Path to an audio file (WAV, FLAC, OGG, etc.).
Returns:
1-D float32 array of shape ``(256,)``, L2-normalized.
"""
mel = self._audio_to_mel(audio_path)
return self._infer(mel)
def encode_batch(self, audio_paths: list[str | Path]) -> np.ndarray:
"""Extract speaker embeddings for multiple audio files.
All mel spectrograms are zero-padded to the longest in the batch
so they can be processed in a single forward pass.
Args:
audio_paths: List of audio file paths.
Returns:
2-D float32 array of shape ``(len(audio_paths), 256)``.
"""
if not audio_paths:
return np.empty((0, 256), dtype=np.float32)
mels = [self._audio_to_mel(p) for p in audio_paths]
# Pad to uniform time length
max_time = max(m.shape[1] for m in mels)
padded = np.zeros((len(mels), mels[0].shape[0], max_time), dtype=np.float32)
for i, m in enumerate(mels):
padded[i, :, : m.shape[1]] = m
return self._infer_batch(padded)
@staticmethod
def similarity(emb1: np.ndarray, emb2: np.ndarray) -> float:
"""Compute cosine similarity between two embeddings.
Both embeddings should already be L2-normalized (as returned by
:meth:`encode`), but this method re-normalizes for safety.
Args:
emb1: 1-D float32 array of shape ``(emb_dim,)``.
emb2: 1-D float32 array of shape ``(emb_dim,)``.
Returns:
Cosine similarity in [-1, 1].
"""
emb1 = emb1.flatten().astype(np.float64)
emb2 = emb2.flatten().astype(np.float64)
norm1 = np.linalg.norm(emb1)
norm2 = np.linalg.norm(emb2)
if norm1 < 1e-12 or norm2 < 1e-12:
return 0.0
return float(np.dot(emb1, emb2) / (norm1 * norm2))
# ------------------------------------------------------------------
# Internal helpers
# ------------------------------------------------------------------
def _audio_to_mel(self, audio_path: str | Path) -> np.ndarray:
"""Load audio and compute mel spectrogram.
Returns:
(n_mels, time) float32 array.
"""
audio = load_audio(audio_path, sr=DEFAULT_SR)
audio = normalize_audio(audio)
mel = compute_mel_spectrogram(
audio,
sr=DEFAULT_SR,
n_fft=DEFAULT_N_FFT,
hop_length=DEFAULT_HOP_LENGTH,
n_mels=DEFAULT_N_MELS,
fmin=DEFAULT_FMIN,
fmax=DEFAULT_FMAX,
)
return mel
def _infer(self, mel: np.ndarray) -> np.ndarray:
"""Run inference on a single mel spectrogram.
Args:
mel: (n_mels, time) float32 array.
Returns:
1-D float32 array of shape ``(emb_dim,)``.
"""
# Add batch dimension: (1, n_mels, time)
mel_batch = mel[np.newaxis, :, :]
return self._infer_batch(mel_batch)[0]
def _infer_batch(self, mel_batch: np.ndarray) -> np.ndarray:
"""Run inference on a batch of mel spectrograms.
Args:
mel_batch: (batch, n_mels, time) float32 array.
Returns:
(batch, emb_dim) float32 array.
"""
if self._mode == "pytorch":
return self._infer_pytorch(mel_batch)
elif self._mode == "onnx":
return self._infer_onnx(mel_batch)
else:
raise RuntimeError(
"SpeakerEncoder not initialised. Use from_pytorch() or from_onnx()."
)
def _infer_pytorch(self, mel_batch: np.ndarray) -> np.ndarray:
"""Run inference with the PyTorch model."""
import torch # noqa: PLC0415
assert self._pytorch_model is not None
tensor = torch.from_numpy(mel_batch).to(self._pytorch_device)
with torch.no_grad():
embedding = self._pytorch_model(tensor)
return embedding.cpu().numpy()
def _infer_onnx(self, mel_batch: np.ndarray) -> np.ndarray:
"""Run inference with the ONNX model."""
assert self._onnx_session is not None
input_name = self._onnx_session.get_inputs()[0].name
output_name = self._onnx_session.get_outputs()[0].name
result = self._onnx_session.run(
[output_name],
{input_name: mel_batch.astype(np.float32)},
)
return result[0]
|