File size: 15,104 Bytes
c3efd49
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
"""Reward function for voice model RL training."""
import torch
import numpy as np
import logging
from typing import Dict, Optional, Tuple

try:
    from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
    import torchaudio
    ASR_AVAILABLE = True
except ImportError:
    ASR_AVAILABLE = False
    logger.warning("ASR dependencies not available. Transcription accuracy will use placeholder.")

logger = logging.getLogger(__name__)


class RewardFunction:
    """
    Computes rewards for voice model outputs based on multiple quality metrics.
    
    Reward components:
    - Clarity: Signal quality and spectral characteristics
    - Naturalness: Prosody and smoothness
    - Accuracy: Similarity to reference (if available)
    """
    
    DEFAULT_PENALTY = -1.0

    def __init__(
        self,
        weights: Optional[Dict[str, float]] = None,
        normalize_range: Tuple[float, float] = (0.0, 1.0),
        use_asr: bool = True,
        asr_model: Optional[str] = "facebook/wav2vec2-base-960h"
    ):
        """
        Initialize reward function.

        Args:
            weights: Component weights {'clarity': 0.33, 'naturalness': 0.33, 'accuracy': 0.34}
            normalize_range: Range for normalized rewards
            use_asr: Whether to use ASR for transcription accuracy
            asr_model: HuggingFace ASR model to use
        """
        if weights is None:
            weights = {
                'clarity': 0.33,
                'naturalness': 0.33,
                'accuracy': 0.34
            }

        # Validate weights
        if not np.isclose(sum(weights.values()), 1.0):
            raise ValueError(f"Weights must sum to 1.0, got {sum(weights.values())}")

        self.weights = weights
        self.normalize_range = normalize_range
        self.use_asr = use_asr and ASR_AVAILABLE

        # Initialize ASR model if requested
        self.asr_model = None
        self.asr_processor = None
        if self.use_asr:
            try:
                self.asr_processor = Wav2Vec2Processor.from_pretrained(asr_model)
                self.asr_model = Wav2Vec2ForCTC.from_pretrained(asr_model)
                self.asr_model.eval()
                logger.info(f"Loaded ASR model: {asr_model}")
            except Exception as e:
                logger.warning(f"Failed to load ASR model: {e}. Using placeholder accuracy.")
                self.use_asr = False

        logger.info(f"Initialized RewardFunction with weights: {weights}, ASR: {self.use_asr}")
    
    def compute_reward(
        self,
        generated_audio: torch.Tensor,
        reference_audio: Optional[torch.Tensor] = None,
        transcription: Optional[str] = None
    ) -> float:
        """
        Compute composite reward for generated audio.
        
        Args:
            generated_audio: Generated audio tensor
            reference_audio: Optional reference audio for comparison
            transcription: Optional expected transcription
        
        Returns:
            Normalized reward score
        """
        try:
            # Convert to numpy for processing
            if isinstance(generated_audio, torch.Tensor):
                generated_audio = generated_audio.detach().cpu().numpy()
            
            if reference_audio is not None and isinstance(reference_audio, torch.Tensor):
                reference_audio = reference_audio.detach().cpu().numpy()
            
            # Compute individual components
            clarity_score = self._compute_clarity(generated_audio)
            naturalness_score = self._compute_naturalness(generated_audio, reference_audio)
            accuracy_score = self._compute_accuracy(generated_audio, reference_audio, transcription)
            
            # Weighted combination
            reward = (
                self.weights['clarity'] * clarity_score +
                self.weights['naturalness'] * naturalness_score +
                self.weights['accuracy'] * accuracy_score
            )
            
            # Normalize to target range
            reward = self._normalize_reward(reward)
            
            return float(reward)
            
        except Exception as e:
            logger.error(f"Error computing reward: {e}")
            return self.DEFAULT_PENALTY
    
    def _compute_clarity(self, audio: np.ndarray) -> float:
        """
        Compute clarity score based on signal quality.
        
        Measures:
        - Signal-to-noise ratio
        - Spectral flatness
        - Absence of clipping
        
        Args:
            audio: Audio waveform
        
        Returns:
            Clarity score in [0, 1]
        """
        score = 0.0
        
        # Check for clipping
        clipping_ratio = np.mean(np.abs(audio) > 0.99)
        clipping_score = 1.0 - clipping_ratio
        score += 0.3 * clipping_score
        
        # Estimate SNR
        signal_power = np.mean(audio ** 2)
        if signal_power > 1e-10:
            # Simple noise estimation from quietest samples
            sorted_power = np.sort(audio ** 2)
            noise_floor = np.mean(sorted_power[:max(1, len(sorted_power) // 20)])
            snr = 10 * np.log10(signal_power / max(noise_floor, 1e-10))
            snr_score = np.clip(snr / 30.0, 0.0, 1.0)  # Normalize to [0, 1]
            score += 0.4 * snr_score
        else:
            score += 0.0
        
        # Spectral flatness (lower is better for speech)
        try:
            fft = np.fft.rfft(audio)
            magnitude = np.abs(fft)
            geometric_mean = np.exp(np.mean(np.log(magnitude + 1e-10)))
            arithmetic_mean = np.mean(magnitude)
            flatness = geometric_mean / (arithmetic_mean + 1e-10)
            flatness_score = 1.0 - flatness  # Invert: lower flatness is better
            score += 0.3 * flatness_score
        except:
            score += 0.15  # Neutral score if computation fails
        
        return np.clip(score, 0.0, 1.0)
    
    def _compute_naturalness(
        self,
        audio: np.ndarray,
        reference: Optional[np.ndarray] = None
    ) -> float:
        """
        Compute naturalness score based on prosody and smoothness.
        
        Measures:
        - Smoothness (absence of abrupt changes)
        - Energy distribution
        - Similarity to reference if available
        
        Args:
            audio: Generated audio
            reference: Optional reference audio
        
        Returns:
            Naturalness score in [0, 1]
        """
        score = 0.0
        
        # Smoothness: penalize abrupt changes
        if len(audio) > 1:
            diff = np.diff(audio)
            smoothness = 1.0 - np.clip(np.std(diff) / 0.1, 0.0, 1.0)
            score += 0.4 * smoothness
        else:
            score += 0.2
        
        # Energy distribution: should not be too uniform or too spiky
        if len(audio) > 10:
            frame_size = len(audio) // 10
            frame_energies = [
                np.mean(audio[i:i+frame_size] ** 2)
                for i in range(0, len(audio) - frame_size, frame_size)
            ]
            energy_std = np.std(frame_energies)
            # Optimal std is around 0.01-0.1
            energy_score = 1.0 - np.clip(abs(energy_std - 0.05) / 0.1, 0.0, 1.0)
            score += 0.3 * energy_score
        else:
            score += 0.15
        
        # Similarity to reference if available
        if reference is not None:
            try:
                # Align lengths
                min_len = min(len(audio), len(reference))
                audio_aligned = audio[:min_len]
                reference_aligned = reference[:min_len]
                
                # Compute correlation
                correlation = np.corrcoef(audio_aligned, reference_aligned)[0, 1]
                correlation_score = (correlation + 1.0) / 2.0  # Map [-1, 1] to [0, 1]
                score += 0.3 * correlation_score
            except:
                score += 0.15
        else:
            score += 0.3  # Neutral score if no reference
        
        return np.clip(score, 0.0, 1.0)
    
    def _compute_accuracy(
        self,
        audio: np.ndarray,
        reference: Optional[np.ndarray] = None,
        transcription: Optional[str] = None
    ) -> float:
        """
        Compute accuracy score based on similarity to reference and/or transcription.

        Args:
            audio: Generated audio
            reference: Optional reference audio
            transcription: Optional expected transcription

        Returns:
            Accuracy score in [0, 1]
        """
        score = 0.0
        num_components = 0

        # Component 1: Audio similarity to reference
        if reference is not None:
            try:
                # Align lengths
                min_len = min(len(audio), len(reference))
                audio_aligned = audio[:min_len]
                reference_aligned = reference[:min_len]

                # Mean squared error (lower is better)
                mse = np.mean((audio_aligned - reference_aligned) ** 2)
                mse_score = np.exp(-mse * 10)  # Exponential decay

                # Correlation
                correlation = np.corrcoef(audio_aligned, reference_aligned)[0, 1]
                correlation_score = (correlation + 1.0) / 2.0

                # Combined audio similarity score
                audio_sim_score = 0.5 * mse_score + 0.5 * correlation_score
                score += audio_sim_score
                num_components += 1

            except Exception as e:
                logger.debug(f"Error computing audio similarity: {e}")

        # Component 2: Transcription accuracy using ASR
        if transcription and self.use_asr and self.asr_model is not None:
            try:
                trans_score = self._compute_transcription_accuracy(audio, transcription)
                score += trans_score
                num_components += 1
            except Exception as e:
                logger.debug(f"Error computing transcription accuracy: {e}")

        # Return average score or neutral if no components
        if num_components > 0:
            return np.clip(score / num_components, 0.0, 1.0)
        else:
            return 0.5

    def _compute_transcription_accuracy(
        self,
        audio: np.ndarray,
        expected_transcription: str,
        sample_rate: int = 16000
    ) -> float:
        """
        Compute transcription accuracy using ASR.

        Args:
            audio: Audio waveform
            expected_transcription: Expected transcription text
            sample_rate: Audio sample rate

        Returns:
            Transcription accuracy score in [0, 1]
        """
        try:
            # Convert to tensor
            audio_tensor = torch.FloatTensor(audio)

            # Resample if needed (ASR models typically use 16kHz)
            if sample_rate != 16000:
                resampler = torchaudio.transforms.Resample(sample_rate, 16000)
                audio_tensor = resampler(audio_tensor)

            # Process audio
            input_values = self.asr_processor(
                audio_tensor,
                sampling_rate=16000,
                return_tensors="pt"
            ).input_values

            # Get transcription
            with torch.no_grad():
                logits = self.asr_model(input_values).logits
                predicted_ids = torch.argmax(logits, dim=-1)
                transcription = self.asr_processor.decode(predicted_ids[0])

            # Compute similarity (simple word error rate approximation)
            score = self._compute_text_similarity(
                transcription.lower().strip(),
                expected_transcription.lower().strip()
            )

            return score

        except Exception as e:
            logger.debug(f"Error in ASR transcription: {e}")
            return 0.5

    def _compute_text_similarity(self, predicted: str, expected: str) -> float:
        """
        Compute text similarity between predicted and expected transcriptions.

        Uses a simple Levenshtein distance-based metric.

        Args:
            predicted: Predicted transcription
            expected: Expected transcription

        Returns:
            Similarity score in [0, 1]
        """
        if not expected:
            return 0.5

        # Simple word-level comparison
        pred_words = set(predicted.split())
        exp_words = set(expected.split())

        if not exp_words:
            return 0.5

        # Jaccard similarity
        intersection = len(pred_words & exp_words)
        union = len(pred_words | exp_words)

        if union == 0:
            return 0.0

        return intersection / union
    
    def _normalize_reward(self, reward: float) -> float:
        """
        Normalize reward to target range.
        
        Args:
            reward: Raw reward value (assumed to be in [0, 1])
        
        Returns:
            Normalized reward
        """
        min_val, max_val = self.normalize_range
        return min_val + (max_val - min_val) * np.clip(reward, 0.0, 1.0)
    
    def get_reward_components(
        self,
        generated_audio: torch.Tensor,
        reference_audio: Optional[torch.Tensor] = None,
        transcription: Optional[str] = None
    ) -> Dict[str, float]:
        """
        Get breakdown of reward components.
        
        Args:
            generated_audio: Generated audio tensor
            reference_audio: Optional reference audio
            transcription: Optional expected transcription
        
        Returns:
            Dictionary with component scores
        """
        try:
            # Convert to numpy
            if isinstance(generated_audio, torch.Tensor):
                generated_audio = generated_audio.detach().cpu().numpy()
            
            if reference_audio is not None and isinstance(reference_audio, torch.Tensor):
                reference_audio = reference_audio.detach().cpu().numpy()
            
            clarity = self._compute_clarity(generated_audio)
            naturalness = self._compute_naturalness(generated_audio, reference_audio)
            accuracy = self._compute_accuracy(generated_audio, reference_audio, transcription)
            
            total = (
                self.weights['clarity'] * clarity +
                self.weights['naturalness'] * naturalness +
                self.weights['accuracy'] * accuracy
            )
            
            return {
                'clarity': clarity,
                'naturalness': naturalness,
                'accuracy': accuracy,
                'total': total,
                'normalized': self._normalize_reward(total)
            }
            
        except Exception as e:
            logger.error(f"Error getting reward components: {e}")
            return {
                'clarity': 0.0,
                'naturalness': 0.0,
                'accuracy': 0.0,
                'total': 0.0,
                'normalized': self.DEFAULT_PENALTY
            }