Upload 2 files

app.py

@@ -0,0 +1,79 @@
+import gradio as gr
+import torch
+from transformers import AutoProcessor, VoxtralForConditionalGeneration
+
+MAX_TOKENS = 32000
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+print(f"*** Device: {device}")
+
+# List models
+dict_models = {'Voxtral-Mini-3B-2507': 'Loren/Voxtral-Mini-3B-2507-dup',
+               'Voxtral-Small-24B-2507': 'Loren/Voxtral-Small-24B-2507-dup'}
+
+# Load models
+list_processor = []
+list_model = []
+for model_name in dict_models.values():
+    list_processor.append(AutoProcessor.from_pretrained(model_name))
+    list_model.append(VoxtralForConditionalGeneration.from_pretrained(model_name,
+                                                                      torch_dtype=torch.bfloat16,
+                                                                      device_map=device))
+# Supported languages
+dict_languages = {"English": "en",
+                  "French": "fr",
+                  "German": "de",
+                  "Spanish": "es",
+                  "Italian": "it",
+                  "Portuguese": "pt",
+                  "Dutch": "nl",
+                  "Hindi": "hi"}
+
+@spaces.GPU
+def process_transcript(audio_path, model_name, language):
+    """Process audio with selected Voxtral model and return the generated response"""
+
+    inputs = processor.apply_transcrition_request(language=language, audio=audio_path, model_id=model_name)
+    inputs = inputs.to(device, dtype=torch.bfloat16)
+
+    outputs = model.generate(**inputs, max_new_tokens=MAX_TOKENS)
+    decoded_outputs = processor.batch_decode(outputs[:, inputs.input_ids.shape[1]:], skip_special_tokens=True)
+
+    return decoded_outputs[0]
+
+
+
+# Define Gradio interface
+with gr.Blocks(title="Transcription") as transcript:
+    gr.Markdown("# Audio Transcription")
+    gr.Markdown("#### Choose the language of the audio and the model, then set an audio file to get its transcription.")
+    gr.Markdown("#### **(Voxtral handles audios up to 30 minutes for transcription)**")
+
+    with gr.Row():
+        with gr.Column():
+            sel_language = gr.Dropdown(
+                choices=list(dict_languages.keys()),
+                value="English",
+                label="Select the language of the audio file:"
+            )
+
+            sel_model = gr.Radio(dict_models.keys(), label="Select the model:")
+
+        with gr.Column():
+            sel_audio = gr.Audio(sources=["microphone", "upload"], type="filepath", label="Upload an audio file or record via microphone:")
+
+            submit_transcript = gr.Button("Extract Transcription", variant="primary")
+
+        with gr.Column():
+            text_transcript = gr.Textbox(label="Generated Response", lines=10)
+
+    submit_transcript.click(
+        fn=process_transcript,
+        inputs=[dict_languages[sel_language], dict_models[sel_model], sel_audio],
+        outputs=text_transcript
+    )
+
+
+# Launch the app
+if __name__ == "__main__":
+    transcript.launch(share=True)

requirements.txt

@@ -0,0 +1,4 @@
+mistral-common
+git+https://github.com/huggingface/transformers
+gradio
+torch