support franch

app.py

@@ -3,32 +3,70 @@ import numpy as np
 import torch
 from datasets import load_dataset
 
-from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline
+from transformers import (
+    SpeechT5ForTextToSpeech,
+    SpeechT5HifiGan,
+    SpeechT5Processor,
+    pipeline,
+    VitsModel,
+    VitsTokenizer,
+)
 
 
 device = "cuda:0" if torch.cuda.is_available() else "cpu"
 
 # load speech translation checkpoint
-asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-base", device=device)
+asr_pipe = pipeline(
+    "automatic-speech-recognition", model="openai/whisper-base", device=device
+)
 
+
+# speecht5
 # load text-to-speech checkpoint and speaker embeddings
-processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
+# processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
+# model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts").to(device)
+# vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
 
-model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts").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)
 
-embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
-speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
+# mms
+model = VitsModel.from_pretrained("Matthijs/mms-tts-fra")
+tokenizer = VitsTokenizer.from_pretrained("Matthijs/mms-tts-fra")
 
 
+# 保持 main 函数 speech_to_speech_translation 不变
+# 并根据需要仅更新 translate 和 synthesise 函数
 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": "translate"})
+    outputs = asr_pipe(
+        audio,
+        max_new_tokens=256,
+        generate_kwargs={"task": "transcribe", "language": "fr"},
+        # generate_kwargs={"task": "transcribe"},
+    )
+    print(outputs)
     return outputs["text"]
 
 
+# speecht5
+# def synthesise(text):
+#     inputs = processor(text=text, return_tensors="pt")
+#     speech = model.generate_speech(
+#         inputs["input_ids"].to(device), speaker_embeddings.to(device), vocoder=vocoder
+#     )
+#     return speech.cpu()
+
+
 def synthesise(text):
-    inputs = processor(text=text, return_tensors="pt")
-    speech = model.generate_speech(inputs["input_ids"].to(device), speaker_embeddings.to(device), vocoder=vocoder)
+    inputs = tokenizer(text, return_tensors="pt")
+    input_ids = inputs["input_ids"]
+
+    with torch.no_grad():
+        outputs = model(input_ids)
+
+    speech = outputs.audio[0]
+
     return speech.cpu()
 
 
@@ -41,7 +79,7 @@ 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
+Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in Chinese. 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:
 
 ![Cascaded STST](https://huggingface.co/datasets/huggingface-course/audio-course-images/resolve/main/s2st_cascaded.png "Diagram of cascaded speech to speech translation")
@@ -69,4 +107,4 @@ file_translate = gr.Interface(
 with demo:
     gr.TabbedInterface([mic_translate, file_translate], ["Microphone", "Audio File"])
 
-demo.launch()
+demo.launch(share=True)

requirements.txt

@@ -1,4 +1,5 @@
 torch
-git+https://github.com/huggingface/transformers
+# git+https://github.com/huggingface/transformers
+git+https://github.com/hollance/transformers.git@6900e8ba6532162a8613d2270ec2286c3f58f57b
 datasets
 sentencepiece