Update app.py

app.py

@@ -1,34 +1,128 @@
-from transformers import AutoTokenizer, VitsModel, set_seed
+# imports
 import gradio as gr
-import torch
+import json
+import librosa
+import os
 import soundfile as sf
+import tempfile
+import uuid
+import torch
+from transformers import AutoTokenizer, VitsModel, set_seed, AutoModelForCausalLM, AutoTokenizer, pipeline
 
+from nemo.collections.asr.models import ASRModel
+from nemo.collections.asr.parts.utils.streaming_utils import FrameBatchMultiTaskAED
+from nemo.collections.asr.parts.utils.transcribe_utils import get_buffered_pred_feat_multitaskAED
 
 
-# Initialize the VITS model, tokenizer, and seed
 
-vits_model = VitsModel.from_pretrained("facebook/mms-tts-eng")
-vits_tokenizer = AutoTokenizer.from_pretrained("facebook/mms-tts-eng")
-set_seed(2020)    
+torch.random.manual_seed(0)
+proc_model_name = "SanctumAI/Phi-3-mini-4k-instruct-GGUF"
+proc_model = AutoModelForCausalLM.from_pretrained(proc_model_name)
+proc_model.to("cpu")
+proc_tokenizer = AutoTokenizer.from_pretrained(proc_model_name)
 
 
-def text_to_speech(text_response):
-    inputs = vits_tokenizer(text=text_response, return_tensors="pt")
-    with torch.no_grad():
-        outputs = vits_model(**inputs)
-    waveform = outputs.waveform[0]
-    sf.write('output.wav', waveform.numpy(), vits_model.config.sampling_rate)
 
-    return 'output.wav'
+
+
+SAMPLE_RATE = 16000 # Hz
+MAX_AUDIO_MINUTES = 10 # wont try to transcribe if longer than this
+
+model = ASRModel.from_pretrained("nvidia/canary-1b")
+model.eval()
+
+# make sure beam size always 1 for consistency
+model.change_decoding_strategy(None)
+decoding_cfg = model.cfg.decoding
+decoding_cfg.beam.beam_size = 1
+model.change_decoding_strategy(decoding_cfg)
+
+
+
+amp_dtype = torch.float16
+
+
+def convert_audio(audio_filepath, tmpdir, utt_id):
+
+	data, sr = librosa.load(audio_filepath, sr=None, mono=True)
+
+	duration = librosa.get_duration(y=data, sr=sr)
+
+	if sr != SAMPLE_RATE:
+		data = librosa.resample(data, orig_sr=sr, target_sr=SAMPLE_RATE)
+
+	out_filename = os.path.join(tmpdir, utt_id + '.wav')
+
+	# save output audio
+	sf.write(out_filename, data, SAMPLE_RATE)
+
+	return out_filename, duration
+
+def transcribe(audio_filepath):
+
+	if audio_filepath is None:
+		raise gr.Error("Please provide some input audio: either upload an audio file or use the microphone")
+	
+	utt_id = uuid.uuid4()
+	with tempfile.TemporaryDirectory() as tmpdir:
+		converted_audio_filepath, duration = convert_audio(audio_filepath, tmpdir, str(utt_id))
+        
+		# make manifest file and save
+		manifest_data = {
+			"audio_filepath": converted_audio_filepath,
+			"source_lang": "en",
+			"target_lang": "en",
+			"taskname": "asr",
+			"pnc": "no",
+			"answer": "predict",
+			"duration": str(duration),
+		}
+
+		manifest_filepath = os.path.join(tmpdir, f'{utt_id}.json')
+
+		with open(manifest_filepath, 'w') as fout:
+			line = json.dumps(manifest_data)
+			fout.write(line + '\n')
+
+		output_text = model.transcribe(manifest_filepath)[0]
+		
+	return output_text
+
+
+
+start = {"role": "system", "content": "You are a helpful digital assistant. Please provide safe, ethical and accurate information to the user."}
+
+
+
+def generate_response(user_input):  
+    messages = [start, {"role": "user", "content": user_input}]
+    inputs = proc_tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")
+    with torch.no_grad():
+        outputs = proc_model.generate(
+            inputs, 
+            max_new_tokens=48,
+        )
+    
+    response = proc_tokenizer.batch_decode(
+        outputs, 
+        skip_special_tokens=True, 
+        clean_up_tokenization_spaces=False,
+    )[0]
+    
+    return response
     
+def CanaryPhi(audio_filepath):
+    user_input = transcribe(audio_filepath)
+    response = generate_response(user_input)
+    return response
 
 
 # Create a Gradio interface
 iface = gr.Interface(
-    fn=text_to_speech, 
-    inputs=gr.Textbox(lines=5, placeholder="Enter your text here..."), 
-    outputs=gr.Audio("response.wav")
+    fn=CanaryPhi, 
+    inputs=gr.Audio(sources="microphone", type="filepath"), 
+    outputs=gr.Textbox(),
 )
 
 # Launch the interface
-iface.launch()
+iface.launch()