fix sampling rates for wavtokenizer

.gitignore

@@ -2,3 +2,4 @@ hw.sol
 inference_wavtokenizer
 .zvuk/**
 inference_wavtokenizer.py
+__pycache__

aulate/metrics_evaluation.py

@@ -152,7 +152,7 @@ def infer_text_to_audio(self, text: str, prompt: str, top_k: int = 50):
 
         # Use default implementation
         max_seq_length = 1024
-        formatted_text = f"Say '{text.upper()}' {prompt}"
+        formatted_text = f"Say '{text.lower()}' {prompt}"
 
         text_tokenized = self.tokenizer(formatted_text, return_tensors="pt")
         text_input_tokens = text_tokenized["input_ids"].to(self.device)
@@ -184,7 +184,7 @@ def calculate_metrics(
         reference_audio: Union[np.ndarray, torch.Tensor],
         generated_audio: Union[np.ndarray, torch.Tensor],
         ref_sr: int = 16000,
-        gen_sr: int = 44100,
+        gen_sr: int = 24000,
     ) -> AudioMetricsResult:
         """Calculate PESQ, STOI, SI-SDR, SIM-O, and SIM-R metrics"""
         # Resample reference and generated audio to a common sample rate for SI-SDR calculation
@@ -397,7 +397,9 @@ def evaluate_batch(
         samples: list,
         prompt: Optional[str] = None,
         batch_metadata: Optional[Dict[str, Any]] = None,
-        save_audio: bool = False
+        save_audio: bool = False,
+        ref_sr: int = 16000,
+        gen_sr: int = 24000,
     ) -> pd.DataFrame:
         """
         Evaluate metrics on a batch of samples.
@@ -427,7 +429,12 @@ def evaluate_batch(
                         generated_audio = generated_audio.audio_data.squeeze()
                     text = None
 
-                metrics = self.calculate_metrics(reference_audio, generated_audio)
+                metrics = self.calculate_metrics(
+                    reference_audio,
+                    generated_audio,
+                    ref_sr=ref_sr,
+                    gen_sr=gen_sr
+                )
                 print({'PESQ': metrics.pesq,
                     'STOI': metrics.stoi,
                     'SI-SDR': metrics.si_sdr,
@@ -453,8 +460,8 @@ def evaluate_batch(
                     audio_filename_ = f"orig_{idx}.wav"
                     audio_path = os.path.join(self.gen_audio_dir, audio_filename)
                     audio_path_ = os.path.join(self.gen_audio_dir, audio_filename_)
-                    sf.write(audio_path, generated_audio.astype(np.float32), 44100)
-                    sf.write(audio_path_, reference_audio.astype(np.float32), 16000)
+                    sf.write(audio_path, generated_audio.detach().cpu().numpy().astype(np.float32), gen_sr)
+                    sf.write(audio_path_, reference_audio.astype(np.float32), ref_sr)
                     result_dict['audio_path'] = audio_path
 
                 results.append(result_dict)
@@ -483,7 +490,11 @@ def evaluate_on_librispeech(
     ):
         """Evaluate metrics on LibriSpeech samples"""
         print(f"Loading LibriSpeech dataset ({subset})...")
-        dataset = torchaudio.datasets.LIBRISPEECH("./data", url=subset, download=True)
+
+        dataset_path = "./data"
+        os.makedirs(dataset_path, exist_ok=True)
+
+        dataset = torchaudio.datasets.LIBRISPEECH(dataset_path, url=subset, download=True)
 
         # Randomly sample entries
         indices = range(600,801)#random.sample(range(len(dataset)), num_samples)
@@ -535,6 +546,6 @@ def evaluate_on_librispeech(
 
     results_df = evaluator.evaluate_on_librispeech(
         num_samples=50,
-        prompt="with a male speaker delivers a very monotone and high-pitched speech with a very fast speed in a setting with almost no noise, creating a clear and loud recording.",
+        prompt="with a voice of male speaker delivers a very monotone and high-pitched speech with a very fast speed in a setting with almost no noise, creating a clear and loud recording.",
     )