use whisper-large-v3 & mms-tts-zlm

.gitignore

@@ -0,0 +1,2 @@
+__pycache__
+.venv

README.md

@@ -9,4 +9,11 @@ app_file: app.py
 pinned: false
 ---
 
+Part of the HuggingFace Audio Processing course.
+
+This is a Gradio wrapper around a (X -> Malay) speech2speech pipeline, where **X** is any language supported by
+`openai/whisper-base`.
+
+The TTS model used is `facebook/mms-tts-zlm`, a pretrained checkpoint for speech in Malay which is part of their **Massively Multilingual Speech** project. The underlying architecture is based on VITS, which generates waveforms directly and does not need a seperate vocoder.
+
 Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

app.py

@@ -1,47 +1,62 @@
+import torch
 import gradio as gr
 import numpy as np
-import torch
-from datasets import load_dataset
 
-from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline
+from transformers import (
+    VitsModel,
+    VitsTokenizer,
+    pipeline,
+)
 
 
 device = "cuda:0" if torch.cuda.is_available() else "cpu"
+torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
+print(f"Using {device} with fp {torch_dtype}")
 
 # load speech translation checkpoint
-asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-base", device=device)
+asr_pipe = pipeline(  # noqa: F821
+    "automatic-speech-recognition",
+    model="openai/whisper-large-v3",
+    device=device,
+    torch_dtype=torch_dtype,
+)
 
-# load text-to-speech checkpoint and speaker embeddings
-processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
+# load text-to-speech checkpoint
 
-model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts").to(device)
-vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
+model = VitsModel.from_pretrained("facebook/mms-tts-zlm")
+tokenizer = VitsTokenizer.from_pretrained("facebook/mms-tts-zlm")
 
-embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
-speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
 
+def synthesise(text):
+    inputs = tokenizer(text=text, return_tensors="pt")
+    input_ids = inputs["input_ids"]
 
-def translate(audio):
-    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate"})
-    return outputs["text"]
+    with torch.no_grad():
+        outputs = model(input_ids)
 
+    speech = outputs["waveform"]
+    return speech
 
-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 translate(audio):
+    outputs = asr_pipe(
+        audio,
+        max_new_tokens=256,
+        generate_kwargs={"task": "transcribe", "language": "ms"},
+    )
+    return outputs["text"]
 
 
 def speech_to_speech_translation(audio):
     translated_text = translate(audio)
     synthesised_speech = synthesise(translated_text)
     synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
-    return 16000, synthesised_speech
+    return 16000, synthesised_speech.T
 
 
 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 **Malay**. 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")
@@ -61,9 +76,10 @@ file_translate = gr.Interface(
     fn=speech_to_speech_translation,
     inputs=gr.Audio(source="upload", type="filepath"),
     outputs=gr.Audio(label="Generated Speech", type="numpy"),
-    examples=[["./example.wav"]],
+    examples="./examples",
     title=title,
     description=description,
+    live=True,
 )
 
 with demo:

requirements.txt

@@ -1,4 +1,4 @@
 torch
-git+https://github.com/huggingface/transformers
 datasets
 sentencepiece
+transformers