Update app.py

app.py

@@ -5,6 +5,35 @@ from datasets import load_dataset
 
 from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline
 
+speaker_embedding_path = "./speaker_embedding.npy"
+
+replacements = [
+    ("&", "og"),
+    ("\r", " "),
+    ("´", ""),
+    ("\\", ""),
+    ("¨", " "),
+    ("Å", "AA"),
+    ("Æ", "AE"),
+    ("É", "E"),
+    ("Ö", "OE"),
+    ("Ø", "OE"),
+    ("á", "a"),
+    ("ä", "ae"),
+    ("å", "aa"),
+    ("è", "e"),
+    ("î", "i"),
+    ("ô", "oe"),
+    ("ö", "oe"),
+    ("ø", "oe"),
+    ("ü", "y"),
+]
+
+
+def replace_danish_letters(text):
+    for src, dst in replacements:
+        text = text.replace(src, dst)
+    return text
 
 device = "cuda:0" if torch.cuda.is_available() else "cpu"
 
@@ -12,17 +41,17 @@ device = "cuda:0" if torch.cuda.is_available() else "cpu"
 asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-base", device=device)
 
 # load text-to-speech checkpoint and speaker embeddings
-processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
+processor = SpeechT5Processor.from_pretrained("JackismyShephard/speecht5_tts-finetuned-nst-da")
 
-model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts").to(device)
+model = SpeechT5ForTextToSpeech.from_pretrained("JackismyShephard/speecht5_tts-finetuned-nst-da").to(device)
 vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
 
-embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
-speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
+speaker_embedding = np.load(speaker_embedding_path)
+speaker_embeddings = torch.tensor(speaker_embedding).unsqueeze(0)
 
 
 def translate(audio):
-    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate"})
+    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "transcribe", "language": "da"})
     return outputs["text"]
 
 
@@ -34,6 +63,8 @@ def synthesise(text):
 
 def speech_to_speech_translation(audio):
     translated_text = translate(audio)
+    translated_text = replace_danish_letters(translated_text)
+    print(translated_text)
     synthesised_speech = synthesise(translated_text)
     synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
     return 16000, synthesised_speech
@@ -42,8 +73,7 @@ def speech_to_speech_translation(audio):
 title = "Cascaded STST"
 description = """
 Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in English. Demo uses OpenAI's [Whisper Base](https://huggingface.co/openai/whisper-base) model for speech translation, and Microsoft's
-[SpeechT5 TTS](https://huggingface.co/microsoft/speecht5_tts) model for text-to-speech:
-
+[SpeechT5 TTS](https://huggingface.co/microsoft/speecht5_tts) model fine-tuned by [JackismyShephard](https://huggingface.co/JackismyShephard) for Danish for text-to-speech:
 ![Cascaded STST](https://huggingface.co/datasets/huggingface-course/audio-course-images/resolve/main/s2st_cascaded.png "Diagram of cascaded speech to speech translation")
 """
 
@@ -51,7 +81,7 @@ demo = gr.Blocks()
 
 mic_translate = gr.Interface(
     fn=speech_to_speech_translation,
-    inputs=gr.Audio(source="microphone", type="filepath"),
+    inputs=gr.Audio(sources=["microphone"], type="filepath"),
     outputs=gr.Audio(label="Generated Speech", type="numpy"),
     title=title,
     description=description,
@@ -59,7 +89,7 @@ mic_translate = gr.Interface(
 
 file_translate = gr.Interface(
     fn=speech_to_speech_translation,
-    inputs=gr.Audio(source="upload", type="filepath"),
+    inputs=gr.Audio(sources=["upload"], type="filepath"),
     outputs=gr.Audio(label="Generated Speech", type="numpy"),
     examples=[["./example.wav"]],
     title=title,
@@ -69,4 +99,4 @@ file_translate = gr.Interface(
 with demo:
     gr.TabbedInterface([mic_translate, file_translate], ["Microphone", "Audio File"])
 
-demo.launch()
+demo.launch()