KEEP Learn: separate Space, English UI, modern design, SVG icons

app.py

@@ -0,0 +1,89 @@
+import os, json, sys, numpy as np, gradio as gr, torch
+from huggingface_hub import snapshot_download
+
+PRIVATE = os.environ.get("MODEL_REPO", "")
+d = snapshot_download(PRIVATE, token=os.environ.get("HF_TOKEN")) if PRIVATE else os.path.dirname(os.path.abspath(__file__))
+sys.path.insert(0, d)
+import keep_audio_core as K
+cfg = json.load(open(os.path.join(d, "config.json"))) if os.path.exists(os.path.join(d, "config.json")) else {}
+R = K.Recognizer(os.path.join(d, "audio2b_KEEP_graine0.pt"), os.path.join(d, "audio2b_normalisation.pt"), gain=cfg.get("gain", "peak"))
+
+CSS = """
+@import url('https://fonts.googleapis.com/css2?family=Cormorant+Garamond:wght@600;700&family=Space+Grotesk:wght@400;500;600&display=swap');
+.gradio-container { max-width: 920px !important; margin: auto !important; background: #F5F4F0 !important; font-family: 'Space Grotesk', system-ui, sans-serif !important; }
+.hero { background: linear-gradient(135deg, #0D1B2A 0%, #1B3A5C 100%); border-radius: 22px; padding: 34px 32px 26px; color: #fff; margin-bottom: 22px; box-shadow: 0 12px 40px rgba(13,27,42,.25); }
+.hero .hero-top { display: flex; gap: 18px; align-items: flex-start; }
+.hero .hero-icon { flex: 0 0 auto; width: 56px; height: 56px; border-radius: 16px; background: rgba(201,162,39,.18); border: 1px solid rgba(201,162,39,.5); display: flex; align-items: center; justify-content: center; color: #C9A227; }
+.hero h1 { font-family: 'Cormorant Garamond', Georgia, serif; font-size: 46px; margin: 0 0 6px; font-weight: 700; letter-spacing: -.5px; line-height: 1.05; }
+.hero .tagline { font-size: 16.5px; opacity: .92; margin: 0 0 16px; line-height: 1.5; }
+.hero .pills { display: flex; flex-wrap: wrap; gap: 8px; }
+.hero .pill { background: rgba(255,255,255,.10); border: 1px solid rgba(255,255,255,.22); border-radius: 999px; padding: 6px 14px; font-size: 13.5px; }
+.hero .pill b { color: #C9A227; }
+button.primary { background: #0D1B2A !important; border-radius: 12px !important; }
+.footnote { color: #64748b; font-size: 13.5px; line-height: 1.55; margin-top: 18px; }
+"""
+
+# Simple, professional mark: a plus sign for "learning something new", drawn as SVG, not an emoji.
+PLUS_SVG = """<svg width="28" height="28" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round"><path d="M12 5v14M5 12h14"/></svg>"""
+
+HERO = f"""<div class="hero"><div class="hero-top"><div class="hero-icon">{PLUS_SVG}</div><div>
+<h1>KEEP Learn</h1>
+<p class="tagline">Teach KEEP words it has <b>never heard</b>, live, with no retraining: it keeps the fingerprint of your examples, then recognizes by resemblance.</p>
+</div></div><div class="pills">
+<span class="pill"><b>2–5</b> words per session</span>
+<span class="pill"><b>1–3</b> examples per word</span>
+<span class="pill">Session only — forgotten on close</span>
+</div></div>"""
+
+def summary(mem):
+    if not mem: return "_No word learned yet._"
+    return "**Learned words:** " + " · ".join(f"{w} ({len(v)} example{'s' if len(v) > 1 else ''})" for w, v in mem.items())
+
+def add_example(name, audio, gain_choice, mem):
+    mem = dict(mem or {}); name = (name or "").strip().lower()
+    if not name: return mem, summary(mem), "Type the word name first.", gr.update()
+    if audio is None: return mem, summary(mem), "Record the word first.", gr.update()
+    R.gain = gain_choice; sr, wav = audio
+    mem[name] = list(mem.get(name, [])) + [R.embedding(wav, sr).tolist()]
+    n = len(mem[name]); tip = " Add 1 or 2 more examples to make it more sure." if n < 3 else ""
+    return mem, summary(mem), f"Example no. {n} of “{name}” memorized.{tip}", None
+
+def forget_all(): return {}, summary({}), "Everything forgotten.", None
+
+def recognize(audio, gain_choice, mem):
+    if not mem or len(mem) < 2: return None, "Teach it at least **2 words** first (field above)."
+    if audio is None: return None, "Say one of the learned words."
+    R.gain = gain_choice; sr, wav = audio; q = R.embedding(wav, sr)
+    names = list(mem); P = torch.nn.functional.normalize(torch.stack([torch.tensor(mem[w]).mean(0) for w in names]), dim=-1)
+    sim = (P @ q).numpy(); p = np.exp(20 * (sim - sim.max())); p = p / p.sum()
+    i = int(sim.argmax()); j = np.argsort(-sim)[1] if len(sim) > 1 else i
+    gap = sim[i] - sim[j]
+    msg = (f"I heard: **{names[i]}**." if gap > 0.05 else f"I'm hesitating… rather **{names[i]}**, but it also sounds like “{names[j]}”. Add examples.")
+    return {w: float(x) for w, x in zip(names, p)}, msg
+
+with gr.Blocks(title="KEEP Learn", css=CSS) as demo:
+    gr.HTML(HERO)
+    gr.Markdown("**1.** Type a name, record the word (1 to 3 times) and click “Memorize”. Do this for 2 to 5 words.\n"
+                "**2.** Below, say one of the learned words again: KEEP tells you which one it recognizes.")
+    gain_radio = gr.Radio(["peak", "none"], value=cfg.get("gain", "peak"), label="Volume setting (peak = automatic, none = raw)")
+    mem = gr.State({})
+    with gr.Row():
+        name = gr.Textbox(label="Word name", placeholder="e.g.: bonjour")
+        ex_audio = gr.Audio(sources=["microphone", "upload"], type="numpy", label="Say this word")
+    with gr.Row():
+        b_add = gr.Button("Memorize this example", variant="primary"); b_forget = gr.Button("Forget everything")
+    info = gr.Markdown(); learned = gr.Markdown(summary({}))
+    b_add.click(add_example, [name, ex_audio, gain_radio, mem], [mem, learned, info, ex_audio])
+    b_forget.click(forget_all, None, [mem, learned, info, ex_audio])
+    gr.Markdown("---")
+    with gr.Row():
+        q_audio = gr.Audio(sources=["microphone", "upload"], type="numpy", label="Say one of the learned words again")
+        q_lab = gr.Label(label="Resemblance to each learned word")
+    q_msg = gr.Markdown()
+    gr.Button("What did I say?", variant="primary").click(recognize, [q_audio, gain_radio, mem], [q_lab, q_msg])
+    gr.Markdown("_Experimental: what it learns here is forgotten when you close the page. "
+                "It works best with the same voice as the examples, and 2 to 3 examples per word._",
+                elem_classes=["footnote"])
+    gr.Markdown("**Limits.** It recognizes word sounds, it does not understand speech. A noisy microphone or words that sound very alike can lower accuracy.",
+                elem_classes=["footnote"])
+demo.launch()