Upload source application

README.md

@@ -1,8 +1,8 @@
 ---
-title: Whisper Tiny Ja Lora Demo
-emoji: 👁
-colorFrom: red
-colorTo: indigo
+title: Whisper Ja Demo
+emoji: 🐨
+colorFrom: yellow
+colorTo: red
 sdk: gradio
 sdk_version: 6.9.0
 app_file: app.py

app.py

@@ -0,0 +1,144 @@
+import gradio as gr
+import torch
+import numpy as np
+from transformers import AutoProcessor, WhisperForConditionalGeneration
+from peft import PeftModel
+
+# ── Constants ──────────────────────────────────────────────────────────────────
+BASE_MODEL_ID = "openai/whisper-tiny"
+LORA_MODEL_ID = "dungca/whisper-tiny-ja-lora"
+SAMPLING_RATE = 16000
+
+# ── Load model (cached after first load) ───────────────────────────────────────
+def load_model():
+    print("Loading model...")
+    processor = AutoProcessor.from_pretrained(BASE_MODEL_ID)
+    base_model = WhisperForConditionalGeneration.from_pretrained(BASE_MODEL_ID)
+    model = PeftModel.from_pretrained(base_model, LORA_MODEL_ID)
+    model.eval()
+    if torch.cuda.is_available():
+        model = model.cuda()
+        print("Using GPU")
+    else:
+        print("Using CPU")
+    return processor, model
+
+processor, model = load_model()
+
+# ── Inference ──────────────────────────────────────────────────────────────────
+def transcribe(audio):
+    if audio is None:
+        return "⚠️ Vui lòng cung cấp audio (upload file hoặc record từ mic)."
+
+    sr, audio_array = audio
+
+    # Convert to float32 mono
+    audio_array = audio_array.astype(np.float32)
+    if audio_array.ndim > 1:
+        audio_array = audio_array.mean(axis=1)
+
+    # Normalize
+    max_val = np.abs(audio_array).max()
+    if max_val > 0:
+        audio_array = audio_array / max_val
+
+    # Resample if needed
+    if sr != SAMPLING_RATE:
+        import librosa
+        audio_array = librosa.resample(audio_array, orig_sr=sr, target_sr=SAMPLING_RATE)
+
+    # Process
+    inputs = processor(
+        audio_array,
+        sampling_rate=SAMPLING_RATE,
+        return_tensors="pt"
+    )
+
+    if torch.cuda.is_available():
+        inputs = {k: v.cuda() for k, v in inputs.items()}
+
+    with torch.no_grad():
+        predicted_ids = model.generate(
+            inputs["input_features"],
+            language="japanese",
+            task="transcribe",
+            max_new_tokens=256,
+        )
+
+    transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
+    return transcription.strip() if transcription.strip() else "（音声が認識できませんでした）"
+
+
+# ── UI ─────────────────────────────────────────────────────────────────────────
+CSS = """
+.gradio-container {
+    max-width: 800px !important;
+    margin: auto !important;
+}
+.title {
+    text-align: center;
+    font-size: 2rem;
+    font-weight: 700;
+    margin-bottom: 0.25rem;
+}
+.subtitle {
+    text-align: center;
+    color: #666;
+    margin-bottom: 1.5rem;
+}
+footer { display: none !important; }
+"""
+
+with gr.Blocks(css=CSS, title="Whisper Japanese ASR Demo") as demo:
+
+    gr.HTML("""
+        <div class="title">🎙️ Whisper Japanese ASR</div>
+        <div class="subtitle">
+            LoRA fine-tuned on <b>ReazonSpeech</b> dataset ·
+            Base: <code>openai/whisper-tiny</code> ·
+            Adapter: <a href="https://huggingface.co/dungca/whisper-tiny-ja-lora" target="_blank">dungca/whisper-tiny-ja-lora</a>
+        </div>
+    """)
+
+    with gr.Tab("🎤 Record từ Mic"):
+        mic_input = gr.Audio(
+            sources=["microphone"],
+            type="numpy",
+            label="Nói tiếng Nhật vào mic...",
+        )
+        mic_btn = gr.Button("📝 Transcribe", variant="primary")
+        mic_output = gr.Textbox(
+            label="Kết quả phiên âm (日本語)",
+            placeholder="Kết quả sẽ hiện ở đây...",
+            lines=3,
+            show_copy_button=True,
+        )
+        mic_btn.click(fn=transcribe, inputs=mic_input, outputs=mic_output)
+
+    with gr.Tab("📁 Upload File"):
+        file_input = gr.Audio(
+            sources=["upload"],
+            type="numpy",
+            label="Upload file audio (wav, mp3, m4a...)",
+        )
+        file_btn = gr.Button("📝 Transcribe", variant="primary")
+        file_output = gr.Textbox(
+            label="Kết quả phiên âm (日本語)",
+            placeholder="Kết quả sẽ hiện ở đây...",
+            lines=3,
+            show_copy_button=True,
+        )
+        file_btn.click(fn=transcribe, inputs=file_input, outputs=file_output)
+
+    gr.HTML("""
+        <div style="margin-top:1.5rem; padding:1rem; background:#f8f9fa; border-radius:8px; font-size:0.9rem; color:#555;">
+            <b>📊 Model Info</b><br>
+            • Base: openai/whisper-tiny (39M params)<br>
+            • Fine-tuning: LoRA (r=16, α=32) trên ReazonSpeech small<br>
+            • CER: 0.525 · Eval loss: 1.177 · Trained on Kaggle P100<br>
+            • <i>whisper-small version đang được train để cải thiện độ chính xác</i>
+        </div>
+    """)
+
+if __name__ == "__main__":
+    demo.launch()

requirements.txt

@@ -0,0 +1,7 @@
+torch>=2.0.0
+transformers>=4.36.0
+peft>=0.18.0
+gradio>=4.0.0
+librosa>=0.10.0
+numpy>=1.24.0
+soundfile>=0.12.0