Upload app(1).py

app(1).py

@@ -0,0 +1,253 @@
+import gradio as gr
+import torch
+import os
+import numpy as np
+import soundfile as sf
+import tempfile
+import json
+import uuid
+from datetime import datetime
+from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
+from peft import PeftModel
+from huggingface_hub import snapshot_download, HfApi, hf_hub_download
+import shutil
+
+# ── Config ────────────────────────────────────────────────────────
+BASE_MODEL    = "openai/whisper-large-v3-turbo"
+ADAPTER_REPO  = "mozed54/whisper-finetuned"
+FEEDBACK_REPO = "mozed54/whisper-feedback"
+HF_TOKEN = os.environ.get("HF_TOKEN")
+
+# ── Chargement modèle ─────────────────────────────────────────────
+device      = "cuda" if torch.cuda.is_available() else "cpu"
+torch_dtype = torch.float16 if device == "cuda" else torch.float32
+
+print(f"Device : {device.upper()}")
+print("Chargement du modele de base...")
+
+base_model = AutoModelForSpeechSeq2Seq.from_pretrained(
+    BASE_MODEL, torch_dtype=torch_dtype, use_safetensors=True,
+)
+
+print("Chargement adapter LoRA...")
+adapter_path = snapshot_download(
+    repo_id=ADAPTER_REPO,
+    repo_type="dataset",
+    token=HF_TOKEN if HF_TOKEN else None,
+)
+model = PeftModel.from_pretrained(base_model, adapter_path)
+model = model.merge_and_unload()
+model.to(device)
+
+processor = AutoProcessor.from_pretrained(BASE_MODEL)
+
+pipe = pipeline(
+    "automatic-speech-recognition",
+    model=model,
+    tokenizer=processor.tokenizer,
+    feature_extractor=processor.feature_extractor,
+    torch_dtype=torch_dtype,
+    device=device,
+    generate_kwargs={"language": "arabic", "task": "transcribe"},
+)
+
+print("Modele pret !")
+
+# ── État global ───────────────────────────────────────────────────
+current_audio = {"path": None, "filename": None}
+
+# ── Fonctions ─────────────────────────────────────────────────────
+def generate_audio_name():
+    """ex: 2026-05-20_10-30-45_a3f9b2.wav"""
+    datetime_str = datetime.utcnow().strftime("%Y-%m-%d_%H-%M-%S")
+    uid          = uuid.uuid4().hex[:6]
+    return f"{datetime_str}_{uid}.wav"
+
+
+def process_audio(audio):
+    if audio is None:
+        return None
+    if isinstance(audio, tuple):
+        sample_rate, data = audio
+        if data.ndim > 1:
+            data = data.mean(axis=1)
+        data = data.astype(np.float32)
+        if np.abs(data).max() > 1.0:
+            data = data / 32768.0
+        tmp = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
+        sf.write(tmp.name, data, sample_rate)
+        return tmp.name
+    return audio
+
+
+def transcribe(audio):
+    path = process_audio(audio)
+    if path is None:
+        return "Aucun audio recu.", gr.update(visible=False)
+    try:
+        result = pipe(path, return_timestamps=False)
+        text   = result["text"].strip()
+        # Sauvegarde le chemin et génère le nom du fichier audio
+        current_audio["path"]     = path
+        current_audio["filename"] = generate_audio_name()
+        return text, gr.update(visible=True)
+    except Exception as e:
+        return f"Erreur : {str(e)}", gr.update(visible=False)
+
+
+def save_feedback(transcription, note):
+    if not transcription or transcription.strip() == "":
+        return "⚠️ Aucune transcription a evaluer."
+    if note is None:
+        return "⚠️ Choisis une note avant de valider."
+
+    try:
+        api = HfApi(token=HF_TOKEN)
+
+        # Crée le repo feedback s'il n'existe pas
+        api.create_repo(
+            repo_id=FEEDBACK_REPO,
+            repo_type="dataset",
+            private=True,
+            exist_ok=True,
+        )
+
+        timestamp    = datetime.utcnow().isoformat()
+        audio_name   = current_audio["filename"] or "inconnu.wav"
+        audio_repo_path = f"audios/{audio_name}"
+
+        # Entrée JSONL
+        entry = {
+            "timestamp":     timestamp,
+            "audio_file":    audio_name,
+            "transcription": transcription,
+            "note":          int(note),
+        }
+
+        # Charge le feedback.jsonl existant ou crée-en un
+        feedback_file = "/tmp/feedback.jsonl"
+        try:
+            existing = hf_hub_download(
+                repo_id=FEEDBACK_REPO,
+                filename="feedback.jsonl",
+                repo_type="dataset",
+                token=HF_TOKEN,
+            )
+            shutil.copy(existing, feedback_file)
+        except Exception:
+            pass
+
+        # Ajoute la ligne
+        with open(feedback_file, "a", encoding="utf-8") as f:
+            f.write(json.dumps(entry, ensure_ascii=False) + "\n")
+
+        # Upload feedback.jsonl
+        api.upload_file(
+            path_or_fileobj=feedback_file,
+            path_in_repo="feedback.jsonl",
+            repo_id=FEEDBACK_REPO,
+            repo_type="dataset",
+            token=HF_TOKEN,
+        )
+
+        # Upload fichier audio dans audios/
+        if current_audio["path"] and os.path.exists(current_audio["path"]):
+            api.upload_file(
+                path_or_fileobj=current_audio["path"],
+                path_in_repo=audio_repo_path,
+                repo_id=FEEDBACK_REPO,
+                repo_type="dataset",
+                token=HF_TOKEN,
+            )
+
+        stars = "⭐" * int(note)
+        return f"✅ Feedback enregistre — {audio_name} — Note : {stars}"
+
+    except Exception as e:
+        return f"❌ Erreur sauvegarde : {str(e)}"
+
+
+# ── CSS ───────────────────────────────────────────────────────────
+css = """
+@import url("https://fonts.googleapis.com/css2?family=DM+Sans:wght@400;500;600&display=swap");
+body, .gradio-container { font-family: "DM Sans", sans-serif !important; background: #0f0f13 !important; }
+.gradio-container { max-width: 860px !important; margin: 0 auto !important; padding: 2rem 1.5rem !important; }
+#app-header { text-align: center; margin-bottom: 2rem; padding: 2rem 0 1.5rem; border-bottom: 1px solid #1e1e2a; }
+#app-header h1 { font-size: 2rem; font-weight: 600; color: #f0f0f5; margin: 0 0 0.4rem; }
+#app-header p { font-size: 0.9rem; color: #6b6b80; margin: 0; }
+.section-label { font-size: 0.72rem; font-weight: 600; color: #4a4a60; letter-spacing: 0.12em; text-transform: uppercase; margin-bottom: 0.6rem; }
+button.primary { background: #5a5aff !important; color: #fff !important; border: none !important; border-radius: 10px !important; font-weight: 500 !important; padding: 0.7rem 1.4rem !important; transition: background 0.2s !important; }
+button.primary:hover { background: #4848e0 !important; }
+button.secondary { background: #1e1e2a !important; color: #a0a0b8 !important; border: 1px solid #2a2a38 !important; border-radius: 10px !important; padding: 0.7rem 1.4rem !important; }
+#transcription-box textarea { background: #0f0f13 !important; border: 1.5px solid #1e1e2a !important; border-radius: 14px !important; color: #e0e0ec !important; font-size: 1rem !important; line-height: 1.7 !important; padding: 1.2rem !important; direction: rtl !important; text-align: right !important; transition: border-color 0.4s, box-shadow 0.4s !important; min-height: 160px !important; }
+#transcription-box.has-text textarea { border-color: #22c97a !important; box-shadow: 0 0 0 3px rgba(34,201,122,0.12) !important; }
+#feedback-box { background: #16161f; border: 1px solid #1e1e2a; border-radius: 14px; padding: 1.2rem 1.5rem; margin-top: 1rem; }
+footer { display: none !important; }
+"""
+
+js = """
+function() {
+    const observer = new MutationObserver(() => {
+        const box = document.querySelector("#transcription-box textarea");
+        const wrapper = document.querySelector("#transcription-box");
+        if (!box || !wrapper) return;
+        if (box.value && box.value.trim().length > 0) { wrapper.classList.add("has-text"); }
+        else { wrapper.classList.remove("has-text"); }
+    });
+    observer.observe(document.body, { childList: true, subtree: true, characterData: true });
+    return [];
+}
+"""
+
+# ── Interface ─────────────────────────────────────────────────────
+with gr.Blocks(title="Whisper Darija", css=css, js=js) as demo:
+
+    gr.HTML('<div id="app-header"><h1>🎙 Whisper · Darija</h1><p>Transcription automatique — dialecte algerien</p></div>')
+
+    gr.HTML('<div class="section-label">Source audio</div>')
+    with gr.Row():
+        with gr.Column(scale=1):
+            mic_input = gr.Audio(
+                sources=["microphone"],
+                type="numpy",
+                label="🎤 Enregistrer ma voix",
+            )
+            mic_btn = gr.Button("Transcrire la voix", variant="primary")
+
+        with gr.Column(scale=1):
+            file_input = gr.Audio(
+                sources=["upload"],
+                type="numpy",
+                label="📂 Charger un fichier audio",
+            )
+            file_btn = gr.Button("Transcrire le fichier", variant="secondary")
+
+    gr.HTML("<hr style='border:none;border-top:1px solid #1e1e2a;margin:1.5rem 0;'>")
+    gr.HTML('<div class="section-label">Transcription</div>')
+
+    output = gr.Textbox(
+        label="",
+        lines=6,
+        placeholder="النص سيظهر هنا",
+        interactive=False,
+        elem_id="transcription-box",
+        show_label=False,
+    )
+
+    # ── Feedback — caché jusqu'à la première transcription ────────
+    with gr.Column(visible=False, elem_id="feedback-box") as feedback_block:
+        gr.HTML('<div class="section-label" style="margin-top:0;">Evaluer la transcription</div>')
+        note_input = gr.Radio(
+            choices=["1", "2", "3", "4", "5"],
+            label="",
+            info="1 = tres mauvais · 5 = parfait",
+            show_label=False,
+        )
+        feedback_btn    = gr.Button("Envoyer le feedback", variant="secondary")
+        feedback_status = gr.Textbox(label="", show_label=False, interactive=False, lines=1)
+
+    mic_btn.click(fn=transcribe, inputs=[mic_input], outputs=[output, feedback_block])
+    file_btn.click(fn=transcribe, inputs=[file_input], outputs=[output, feedback_block])
+    feedback_btn.click(fn=save_feedback, inputs=[output, note_input], outputs=[feedback_status])
+
+demo.launch()