Update app.py

app.py

@@ -3,30 +3,20 @@ import soundfile as sf
 import numpy as np
 import torch, torchaudio
 from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
-from datasets import load_dataset, Audio
-import matplotlib.pyplot as plt
 
-
-MODEL_NAME="carlosdanielhernandezmena/wav2vec2-large-xlsr-53-icelandic-ep10-1000h"
+MODEL_IS="carlosdanielhernandezmena/wav2vec2-large-xlsr-53-icelandic-ep10-1000h"
+MODEL_FO="carlosdanielhernandezmena/wav2vec2-large-xlsr-53-faroese-100h"
 
 torch.random.manual_seed(0)
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 
-model = Wav2Vec2ForCTC.from_pretrained(MODEL_NAME).to(device)
-processor = Wav2Vec2Processor.from_pretrained(MODEL_NAME) # do i need this? can't remember
-
-
-#ds = load_dataset("language-and-voice-lab/samromur_asr",split='train',streaming=True)
-#ds = load_dataset("language-and-voice-lab/samromur_asr",split='test')
-#ds = ds.cast_column("audio", Audio(sampling_rate=16_000))
-
-def show_ex(exnum):
-    #return(ds['audio_id'][exnum])
-    return(exnum)
-    
+model_is = Wav2Vec2ForCTC.from_pretrained(MODEL_IS).to(device)
+processor_is = Wav2Vec2Processor.from_pretrained(MODEL_IS) 
+model_fo = Wav2Vec2ForCTC.from_pretrained(MODEL_FO).to(device)
+processor_fo = Wav2Vec2Processor.from_pretrained(MODEL_FO) 
 
-def recc(a_f):
-    wav, sr = sf.read(a_f, dtype=np.float32)
+def recc(audio_file,model,processor):
+    wav, sr = sf.read(audio_file, dtype=np.float32)
     if len(wav.shape) == 2:
         wav = wav.mean(1)
     if sr != 16000:
@@ -34,31 +24,42 @@ def recc(a_f):
         wav = signal.resample(wav, wlen)
         
     with torch.inference_mode():
-        #wav = torch.from_numpy(wav).unsqueeze(0)
-        #if torch.cuda.is_available():
-        #    wav = wav.cuda()
         input_values = processor(wav,sampling_rate=16000).input_values[0]
         input_values = torch.tensor(input_values, device=device).unsqueeze(0)
         logits = model(input_values).logits
         pred_ids = torch.argmax(logits, dim=-1)
-        #pred_ids= pred_ids[0].cpu().detach()
         xcp = processor.batch_decode(pred_ids)
         return xcp
 
 
 bl = gr.Blocks()
 with bl:
-    audio_file = gr.Audio(type="filepath")
-    text_button = gr.Button("Recognise")
-    text_output = gr.Textbox()
-    
-    text_button.click(recc, inputs=audio_file, outputs=text_output)
 
+    gr.Markdown(
+        """
+    # W2V2 speech recognition
+    Upload a file for recognition with 
+    https://huggingface.co/carlosdanielhernandezmena/wav2vec2-large-xlsr-53-icelandic-ep10-1000h 
+    or https://huggingface.co/carlosdanielhernandezmena/wav2vec2-large-xlsr-53-faroese-100h  
 
-bl.launch()
+    - For some reason, the huggingface 'Hosted inference API' on the model page does not work, but this demo does.  
+    - There is no language model (yet), so it can generate non-words.
+    """
+    )
+
+    with gr.Tabs():
+        with gr.TabItem("Icelandic"):
+            with gr.Row():
+                audio_file = gr.Audio(type="filepath")
+                text_output = gr.Textbox()
+            text_button = gr.Button("Recognise")
+            text_button.click(recc, inputs=[audio_file,model_is,processor_is], outputs=text_output)
+        with gr.TabItem("Faroese"):
+            with gr.Row():
+                audio_file = gr.Audio(type="filepath")
+                text_output = gr.Textbox()
+            text_button = gr.Button("Recognise")
+            text_button.click(recc, inputs=[audio_file,model_fo,processor_fo], outputs=text_output)
 
-#https://mercury-docs.readthedocs.io/en/latest/deploy/hugging-face-spaces/
-#https://huggingface.co/spaces/pplonski/deploy-mercury
-#https://discuss.huggingface.co/t/deploy-interactive-jupyter-notebook-on-spaces-with-mercury/17000
-#https://huggingface.co/docs/transformers/notebooks
+bl.launch()