File size: 11,305 Bytes
17dd01d
 
 
 
 
 
cc40a1e
17dd01d
 
cc40a1e
17dd01d
 
cc40a1e
 
17dd01d
 
 
 
 
 
 
 
 
 
cc40a1e
 
 
 
 
 
 
 
17dd01d
 
 
 
 
 
 
 
 
 
 
 
 
cc40a1e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
17dd01d
 
cc40a1e
 
17dd01d
 
 
 
 
cc40a1e
17dd01d
 
cc40a1e
 
17dd01d
cc40a1e
17dd01d
 
 
 
 
 
 
 
cc40a1e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
17dd01d
 
cc40a1e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
17dd01d
 
cc40a1e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
17dd01d
 
 
 
 
 
 
 
 
 
 
cc40a1e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from __future__ import annotations

import json
import sys
from dataclasses import dataclass
from pathlib import Path
from typing import Iterable

import librosa
import matplotlib.pyplot as plt
import numpy as np
import soundfile as sf
from scipy.optimize import linear_sum_assignment
from scipy.signal import medfilt

ROOT = Path(__file__).resolve().parent
for path in (ROOT, ROOT / "datasets"):
    path_str = str(path)
    if path_str not in sys.path:
        sys.path.insert(0, path_str)

import feature as ls_feature  # noqa: E402


@dataclass
class InferenceResult:
    logits: np.ndarray
    probabilities: np.ndarray
    full_logits: np.ndarray
    full_probabilities: np.ndarray
    frame_hz: float
    duration_seconds: float


def ensure_mono(audio: np.ndarray) -> np.ndarray:
    if audio.ndim == 1:
        return audio.astype(np.float32, copy=False)
    return audio.mean(axis=1, dtype=np.float32)


def load_audio(audio_path: Path) -> tuple[np.ndarray, int]:
    audio, sample_rate = sf.read(audio_path)
    return ensure_mono(audio), sample_rate


def config_from_metadata(metadata: dict) -> dict:
    return {
        "data": {
            "feat": {
                "sample_rate": int(metadata["sample_rate"]),
                "win_length": int(metadata["win_length"]),
                "hop_length": int(metadata["hop_length"]),
                "n_fft": int(metadata["n_fft"]),
                "n_mels": int(metadata["n_mels"]),
            },
            "context_recp": int(metadata["context_recp"]),
            "subsampling": int(metadata["subsampling"]),
            "feat_type": str(metadata["feat_type"]),
            "max_speakers": int(metadata.get("max_speakers", int(metadata["max_nspks"]) - 2)),
        },
        "model": {
            "params": {
                "conv_delay": int(metadata["conv_delay"]),
            }
        },
    }


def extract_features(audio: np.ndarray, sample_rate: int, config: dict) -> np.ndarray:
    target_sr = int(config["data"]["feat"]["sample_rate"])
    if sample_rate != target_sr:
        audio = librosa.resample(audio, orig_sr=sample_rate, target_sr=target_sr)
    frame_shift = int(config["data"]["feat"]["hop_length"])
    subsampling = int(config["data"]["subsampling"])
    usable_samples = (len(audio) // (frame_shift * subsampling)) * (frame_shift * subsampling)
    if usable_samples > 0:
        audio = audio[:usable_samples]
    stft = ls_feature.stft(
        audio,
        frame_size=int(config["data"]["feat"]["win_length"]),
        frame_shift=frame_shift,
    )
    feats = ls_feature.transform(stft, str(config["data"]["feat_type"]))
    feats = ls_feature.splice(feats, int(config["data"]["context_recp"]))
    feats, _ = ls_feature.subsample(feats, feats, subsampling)
    return np.array(feats, copy=True).astype(np.float32, copy=False)


def frame_hz(config: dict) -> float:
    return config["data"]["feat"]["sample_rate"] / (
        config["data"]["feat"]["hop_length"] * config["data"]["subsampling"]
    )


def parse_rttm(rttm_path: Path) -> tuple[list[dict], list[str]]:
    entries = []
    speaker_order = []
    with open(rttm_path, "r", encoding="utf-8") as handle:
        for line in handle:
            parts = line.strip().split()
            if not parts:
                continue
            speaker = parts[7]
            if speaker not in speaker_order:
                speaker_order.append(speaker)
            entries.append(
                {
                    "recording_id": parts[1],
                    "start": float(parts[3]),
                    "duration": float(parts[4]),
                    "speaker": speaker,
                }
            )
    return entries, speaker_order


def rttm_to_frame_matrix(entries: list[dict], speakers: list[str], num_frames: int, frame_rate: float) -> np.ndarray:
    matrix = np.zeros((num_frames, len(speakers)), dtype=np.float32)
    speaker_to_index = {speaker: index for index, speaker in enumerate(speakers)}
    for entry in entries:
        start = int(round(entry["start"] * frame_rate))
        stop = int(round((entry["start"] + entry["duration"]) * frame_rate))
        matrix[start : min(stop, num_frames), speaker_to_index[entry["speaker"]]] = 1.0
    return matrix


def collar_mask(reference: np.ndarray, collar_frames: int) -> np.ndarray:
    if collar_frames <= 0:
        return np.ones(reference.shape[0], dtype=bool)
    mask = np.ones(reference.shape[0], dtype=bool)
    for column in range(reference.shape[1]):
        padded = np.pad(reference[:, column], (1, 1), constant_values=0)
        changes = np.where(np.diff(padded) != 0)[0]
        for change in changes:
            start = max(0, change - collar_frames)
            stop = min(reference.shape[0], change + collar_frames)
            mask[start:stop] = False
    return mask


def _pair_cost(pred_column: np.ndarray, ref_column: np.ndarray) -> float:
    pred_column = pred_column.astype(bool)
    ref_column = ref_column.astype(bool)
    n_ref = ref_column.sum()
    n_sys = pred_column.sum()
    n_map = np.logical_and(pred_column, ref_column).sum()
    miss = max(n_ref - n_sys, 0)
    false_alarm = max(n_sys - n_ref, 0)
    speaker_error = min(n_ref, n_sys) - n_map
    return float(miss + false_alarm + speaker_error)


def map_predictions(
    prediction_binary: np.ndarray,
    reference_binary: np.ndarray,
    valid_mask: np.ndarray,
) -> tuple[np.ndarray, dict[int, int], list[int]]:
    masked_pred = prediction_binary[valid_mask]
    masked_ref = reference_binary[valid_mask]
    num_pred = prediction_binary.shape[1]
    num_ref = reference_binary.shape[1]
    mapped = np.zeros((prediction_binary.shape[0], num_ref), dtype=np.float32)
    assignment: dict[int, int] = {}
    if num_pred == 0 or num_ref == 0:
        return mapped, assignment, list(range(num_pred))
    cost = np.zeros((num_pred, num_ref), dtype=np.float32)
    for pred_index in range(num_pred):
        for ref_index in range(num_ref):
            cost[pred_index, ref_index] = _pair_cost(masked_pred[:, pred_index], masked_ref[:, ref_index])
    row_index, col_index = linear_sum_assignment(cost)
    matched_pred = set()
    for pred_index, ref_index in zip(row_index, col_index):
        mapped[:, ref_index] = prediction_binary[:, pred_index]
        assignment[int(ref_index)] = int(pred_index)
        matched_pred.add(int(pred_index))
    unmatched_pred = [pred_index for pred_index in range(num_pred) if pred_index not in matched_pred]
    return mapped, assignment, unmatched_pred


def compute_der(
    probabilities: np.ndarray,
    reference_binary: np.ndarray,
    threshold: float,
    median_width: int,
    collar_seconds: float,
    frame_rate: float,
) -> dict:
    prediction_binary = (probabilities > threshold).astype(np.float32)
    if median_width > 1:
        prediction_binary = medfilt(prediction_binary, kernel_size=(median_width, 1)).astype(np.float32)
    valid_mask = collar_mask(reference_binary, int(round(collar_seconds * frame_rate)))
    mapped_binary, assignment, unmatched_pred = map_predictions(prediction_binary, reference_binary, valid_mask)
    mapped_probabilities = np.zeros((probabilities.shape[0], reference_binary.shape[1]), dtype=np.float32)
    for ref_index, pred_index in assignment.items():
        mapped_probabilities[:, ref_index] = probabilities[:, pred_index]
    extra_binary = prediction_binary[:, unmatched_pred] if unmatched_pred else np.zeros((prediction_binary.shape[0], 0), dtype=np.float32)
    scored_reference = np.concatenate(
        [reference_binary, np.zeros((reference_binary.shape[0], extra_binary.shape[1]), dtype=np.float32)],
        axis=1,
    )
    scored_prediction = np.concatenate([mapped_binary, extra_binary], axis=1)
    masked_ref = scored_reference[valid_mask]
    masked_pred = scored_prediction[valid_mask]
    n_ref = masked_ref.sum(axis=1)
    n_sys = masked_pred.sum(axis=1)
    miss = np.maximum(n_ref - n_sys, 0).sum()
    false_alarm = np.maximum(n_sys - n_ref, 0).sum()
    mapped_overlap = np.logical_and(masked_ref == 1, masked_pred == 1).sum(axis=1)
    speaker_error = (np.minimum(n_ref, n_sys) - mapped_overlap).sum()
    speaker_scored = masked_ref.sum()
    der = float((miss + false_alarm + speaker_error) / speaker_scored) if speaker_scored else 0.0
    return {
        "der": der,
        "speaker_scored": float(speaker_scored),
        "speaker_miss": float(miss),
        "speaker_false_alarm": float(false_alarm),
        "speaker_error": float(speaker_error),
        "threshold": threshold,
        "median_width": median_width,
        "collar_seconds": collar_seconds,
        "mapped_binary": mapped_binary,
        "mapped_probabilities": mapped_probabilities,
        "valid_mask": valid_mask,
        "assignment": assignment,
        "unmatched_prediction_indices": unmatched_pred,
    }


def write_rttm(
    recording_id: str,
    binary_prediction: np.ndarray,
    output_path: Path,
    frame_rate: float,
    speaker_labels: Iterable[str] | None = None,
) -> None:
    speaker_labels = list(speaker_labels or [f"spk{index:02d}" for index in range(binary_prediction.shape[1])])
    with open(output_path, "w", encoding="utf-8") as handle:
        for speaker_index, speaker in enumerate(speaker_labels):
            padded = np.pad(binary_prediction[:, speaker_index], (1, 1), constant_values=0)
            changes = np.where(np.diff(padded) != 0)[0]
            for start, stop in zip(changes[::2], changes[1::2]):
                start_seconds = start / frame_rate
                duration_seconds = (stop - start) / frame_rate
                handle.write(
                    f"SPEAKER {recording_id} 1 {start_seconds:.3f} {duration_seconds:.3f} <NA> <NA> {speaker} <NA> <NA>\n"
                )


def save_heatmap(
    reference_binary: np.ndarray,
    mapped_binary: np.ndarray,
    mapped_probabilities: np.ndarray,
    frame_rate: float,
    speaker_labels: list[str],
    output_path: Path,
) -> None:
    duration_seconds = reference_binary.shape[0] / frame_rate
    fig, axes = plt.subplots(3, 1, figsize=(16, 8), sharex=True, constrained_layout=True)
    plots = [
        (reference_binary, "Expected (RTTM)", "Greys"),
        (mapped_binary, "Predicted (Mapped, Binary)", "Greys"),
        (mapped_probabilities, "Predicted (Mapped, Probability)", "viridis"),
    ]
    for axis, (matrix, title, cmap) in zip(axes, plots):
        image = axis.imshow(
            matrix.T,
            aspect="auto",
            origin="lower",
            interpolation="nearest",
            extent=[0.0, duration_seconds, -0.5, matrix.shape[1] - 0.5],
            cmap=cmap,
            vmin=0.0,
            vmax=1.0,
        )
        axis.set_title(title)
        axis.set_yticks(range(len(speaker_labels)))
        axis.set_yticklabels(speaker_labels)
        if cmap != "Greys":
            fig.colorbar(image, ax=axis, fraction=0.02, pad=0.01)
    axes[-1].set_xlabel("Time (seconds)")
    output_path.parent.mkdir(parents=True, exist_ok=True)
    fig.savefig(output_path, dpi=200)
    plt.close(fig)


def save_json(payload: dict, output_path: Path) -> None:
    output_path.parent.mkdir(parents=True, exist_ok=True)
    with open(output_path, "w", encoding="utf-8") as handle:
        json.dump(payload, handle, indent=2, sort_keys=True)