Update requirements.txt

app.py

@@ -1,22 +1,25 @@
-import transformers.utils
-
-# Patch missing helper in modern transformers for Optimum compatibility
-if not hasattr(transformers.utils, "is_tf_available"):
-    transformers.utils.is_tf_available = lambda: False
-
 import gradio as gr
 import librosa
 import numpy as np
-import torch
-from optimum.onnxruntime import ORTModelForCTC
+import onnxruntime as ort
+from huggingface_hub import hf_hub_download
 from transformers import AutoProcessor
 
-
 MODEL_ID = "njand/wav2vec2-xls-r-latin"
 FILE_NAME = "model_quantized.onnx"
 
+print("Downloading processor and ONNX model...")
 processor = AutoProcessor.from_pretrained(MODEL_ID)
-model = ORTModelForCTC.from_pretrained(MODEL_ID, file_name=FILE_NAME)
+
+# Download ONNX quantized weights directly from HF Hub
+model_path = hf_hub_download(repo_id=MODEL_ID, filename=FILE_NAME)
+
+# Explicitly initialize ONNX Runtime on CPU to avoid ZeroGPU driver clashes
+session = ort.InferenceSession(
+    model_path, providers=["CPUExecutionProvider"]
+)
+input_name = session.get_inputs()[0].name
+print("ONNX Model loaded successfully!")
 
 
 def transcribe_audio(audio_tuple) -> str:
@@ -26,28 +29,34 @@ def transcribe_audio(audio_tuple) -> str:
     try:
         sample_rate, audio_data = audio_tuple
 
+        # Convert stereo to mono if audio has multiple channels
         if audio_data.ndim > 1:
             audio_data = audio_data.mean(axis=1)
 
+        # Normalize integer PCM data to float32 (-1.0 to 1.0)
         if np.issubdtype(audio_data.dtype, np.integer):
             max_val = np.iinfo(audio_data.dtype).max
             audio_data = audio_data.astype(np.float32) / max_val
         elif audio_data.dtype != np.float32:
             audio_data = audio_data.astype(np.float32)
 
+        # Resample to 16 kHz (required by Wav2Vec2)
         if sample_rate != 16000:
             audio_data = librosa.resample(
                 audio_data, orig_sr=sample_rate, target_sr=16000
             )
 
-        input_values = processor(
-            audio_data, sampling_rate=16000, return_tensors="pt"
+        # Preprocess input audio into NumPy arrays
+        inputs = processor(
+            audio_data, sampling_rate=16000, return_tensors="np"
         ).input_values
 
-        with torch.no_grad():
-            logits = model(input_values).logits
+        # Run ONNX inference
+        onnx_outputs = session.run(None, {input_name: inputs})
+        logits = onnx_outputs[0]
 
-        predicted_ids = torch.argmax(logits, dim=-1)
+        # Decode token predictions
+        predicted_ids = np.argmax(logits, axis=-1)
         transcription = processor.batch_decode(predicted_ids)[0]
 
         return (
@@ -60,6 +69,7 @@ def transcribe_audio(audio_tuple) -> str:
         return f"❌ An error occurred during transcription: {str(e)}"
 
 
+# Build Gradio UI
 with gr.Blocks(theme=gr.themes.Soft(), title="Latin Speech-to-Text") as demo:
     gr.Markdown("# 🎙️ Latin Speech Recognition Demo")
 

requirements.txt

@@ -1,6 +1,4 @@
-optimum[onnxruntime]
-transformers
+onnxruntime
 librosa
-torch==2.11.0
-numpy
-numba>=0.60.0
+transformers
+huggingface-hub