feat: add infer, utils and change app.py for gradio

app.py

@@ -1,21 +1,28 @@
-from transformers import pipeline
+from src import infer, utils
 import gradio as gr
 
-p = pipeline("automatic-speech-recognition")
 
-def transcribe(audio, state=""):
-    text = p(audio)["text"]
-    state += text + " "
-    return state, state
+audio_examples = [
+    [None, "assets/audio/male-indonesian.wav", None],
+    [None, "assets/audio/female-indonesian.wav", None],
+    [None, "assets/audio/male-english.wav", None],
+    [None, "assets/audio/female-english.wav", None],   
+]
 
-gr.Interface(
-    fn=transcribe, 
+demo = gr.Interface(
+    fn=infer.predict, 
     inputs=[
-        gr.Audio(source="microphone", type="filepath", streaming=True), 
-        "state"
+        gr.Radio(label="Language", 
+                 choices=["indonesian","english"], 
+                 value="indonesian"),
+        gr.Audio(label="Speak", source="microphone", type="numpy"),
+        gr.Audio(label="Upload audio", source="upload", type="numpy"),
     ],
-    outputs=[
-        "textbox",
-        "state"
-    ],
-    live=True).launch()
+    outputs=[gr.TextArea(label="Output Text"),],
+    title="OpenAI Whisper Base",
+    description=utils.parsing_text("assets/descriptions.md"),
+    article=""utils.parsing_text("assets/articles.md"),
+    examples=audio_examples,
+)
+
+demo.launch()

assets/articles.md

@@ -0,0 +1 @@
+articles file

assets/audio/readme.md

@@ -0,0 +1,6 @@
+---
+noteId: "b807e330385e11eeb3554fca5e7f7a03"
+tags: []
+
+---
+

assets/descriptions.md

@@ -0,0 +1 @@
+description files

src/infer.py

@@ -0,0 +1,28 @@
+
+from typing import *
+from src import utils
+from transformers import WhisperProcessor, WhisperForConditionalGeneration
+
+model_name: str  = f"openai/whisper-base"
+processor: Any = WhisperProcessor.from_pretrained(model_name)
+model: Any = WhisperForConditionalGeneration.from_pretrained(model_name)
+
+sample_rate: int = 16000
+float_factor: float = 32678.0
+
+
+def predict(language, mic_audio=None, audio=None):
+    if mic_audio is not None:
+        sampling_rate, waveform = mic_audio
+    elif audio is not None:
+        sampling_rate, waveform = audio
+    else:
+        return "(please provide audio)"
+    
+    forced_decoder_ids = processor.get_decoder_prompt_ids(language=language, task="transcribe")
+    
+    waveform = utils.preprocess_audio(sampling_rate, waveform)
+    inputs = processor(audio=waveform, sampling_rate=sample_rate, return_tensors="pt")
+    predicted_ids = model.generate(**inputs, max_length=400, forced_decoder_ids=forced_decoder_ids)
+    transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)
+    return transcription[0]

src/utils.py

@@ -0,0 +1,29 @@
+
+import librosa 
+import torch
+from pathlib import Path
+
+sample_rate: int = 16000
+float_factor: float = 32678.0
+
+def preprocess_audio(sampling_rate, waveform):
+    waveform = waveform / float_factor
+    
+    if len(waveform) > 1:
+        waveform = librosa.to_mono(waveform.T)
+    
+    if sampling_rate != sample_rate:
+        waveform = librosa.resample(waveform, orig_sr=sampling_rate, target_sr=sample_rate)
+    
+    waveform = waveform[:sample_rate*30]
+    waveform = torch.tensor(waveform)
+    return waveform
+
+
+def parsing_text(filepath: str):
+    path = Path(filepath)
+    if path.suffix.lower() not in ('.txt', '.md'):
+        raise ValueError("Invalid file type. Only '.txt' and '.md' files are supported.")
+
+    return path.read_text()
+