Update README.md

README.md

@@ -59,4 +59,123 @@ Dieser kombinierte deutsche Sprachdatensatz kann für verschiedene Zwecke verwen
 - NLP (Natural Language Processing) Forschung 
 - Text-to-Speech Anwendungen 
 
-Bitte beachten Sie jedoch bei der Verwendung dieses Datensatzes die Lizenzbedingungen der einzelnen Quellen sowie etwaige Einschränkungen oder Richtlinien bezüglich des Datenschutzes oder Urheberrechts.
+Bitte beachten Sie jedoch bei der Verwendung dieses Datensatzes die Lizenzbedingungen der einzelnen Quellen sowie etwaige Einschränkungen oder Richtlinien bezüglich des Datenschutzes oder Urheberrechts.
+
+
+
+
+Evaluation results:
+
+| Modell | Word error rate |
+| ----- | -----  |
+| openai/whisper-large-v3-turbo | 14.14% |
+
+openai/whisper-large-v3-turbo
+
+
+Evaluation script
+
+```python
+import datasets
+from transformers import (
+    AutoConfig,
+    AutoModelForSpeechSeq2Seq,
+    AutoProcessor,
+)
+from jiwer import wer, cer, wer_standardize_contiguous
+from tqdm.auto import tqdm
+import torch
+
+BASE_MODEL = "Model Path"
+DATASET_NAME = "asr-german-mixed"
+batch_size=4
+
+cv_data = (
+    datasets.load_from_disk(DATASET_NAME)
+    .cast_column(
+        "audio",
+        datasets.Audio(sampling_rate=16000, decode=True),
+    )
+    .with_format("torch")
+)
+
+
+def get_model(model_name: str):
+    kwargs = {}
+    processor = AutoProcessor.from_pretrained(
+        model_name,
+        legacy=False,
+        trust_remote_code=True,
+    )
+
+    kwargs["attn_implementation"] = "sdpa"
+
+    model = AutoModelForSpeechSeq2Seq.from_pretrained(
+        model_name,
+        **kwargs,
+    )
+
+    model = model.to(torch.bfloat16).cuda()
+    
+    return model, processor
+
+
+model, processor = get_model(model_name=BASE_MODEL)
+model.generation_config.do_sample=False
+model.generation_config.num_beams=1
+
+references = []
+predictions = []
+test_data = cv_data["test"]
+
+for i in tqdm(range(0, len(test_data), batch_size)):
+    batch = test_data[i:i + batch_size]
+
+    audios = []
+    input_features = []
+    for x in batch["audio"]:
+        # Extract the audio features from the audio
+        extracted = processor(
+            audio=x["array"],
+            sampling_rate=16000,
+            return_tensors="pt",
+            return_attention_mask = True,
+        )
+
+        # check if feature extractor return input_features or input_values
+        ft = (
+            "input_values"
+            if hasattr(extracted, "input_values")
+            else "input_features"
+        )
+
+        # append to input_features
+        input_features.append(
+            {
+                ft: getattr(
+                    extracted,
+                    ft,
+                )[0],
+                "attention_mask": extracted.attention_mask
+            }
+        )
+        
+        
+    batch_pad = processor.feature_extractor.pad(
+        input_features,
+        padding="longest",
+        return_tensors="pt",
+    )
+    
+    with torch.no_grad():
+        generated_ids = model.generate(input_features=batch_pad.input_features.to(torch.bfloat16).cuda(), attention_mask = batch_pad.attention_mask.to(torch.bfloat16).cuda(), max_new_tokens=384, language='de')
+    
+    transcriptions = processor.batch_decode(generated_ids, skip_special_tokens=True)
+    references.extend(batch["transkription"])
+    predictions.extend(transcriptions)
+    if i % 1000 == 0:
+        print("WER: ", wer(references, predictions, reference_transform=wer_standardize_contiguous, hypothesis_transform=wer_standardize_contiguous)*100)
+print("*"*20)
+print("WER: ", wer(references, predictions, reference_transform=wer_standardize_contiguous, hypothesis_transform=wer_standardize_contiguous)*100)
+print("*"*20)
+```