Upload folder using huggingface_hub

README.md

@@ -8,7 +8,7 @@ sdk_version: 5.49.1
 app_file: app.py
 pinned: false
 license: apache-2.0
-hardware: cpu-basic
+suggested_hardware: zero-a10g
 tags:
   - track:backyard
   - sponsor:modal
@@ -37,7 +37,7 @@ against 9.4% for the vanilla model.
 - **Base:** `openai/whisper-large-v3-turbo`
 - **Adapter:** [`pradachan/whisper-large-v3-turbo-disfluency-lora`](https://huggingface.co/pradachan/whisper-large-v3-turbo-disfluency-lora)
 - **Trained on:** [Modal](https://modal.com) serverless GPUs (LoRA fine-tuning and the probe/eval harness both ran on Modal).
-- **Hardware:** CPU (upgrade to t4-small or ZeroGPU for live transcription)
+- **Hardware:** ZeroGPU (`@spaces.GPU`) — live transcription runs on an on-demand A10G.
 
 The gallery examples are real results from the DisfluencySpeech test set, plus one
 honest limitation. Vanilla Whisper already deletes most filler "um/uh" natively, so

app.py

@@ -49,7 +49,14 @@ _MODELS = {}
 
 
 def load_models():
-    """Load processor + both models once, on first GPU call. Returns a dict."""
+    """Load the processor and one base model + LoRA once, on first GPU call.
+
+    We deliberately load the base weights a *single* time and attach the LoRA
+    on top. Vanilla decoding runs inside a ``disable_adapter()`` context, so
+    there is only one copy of whisper-large-v3-turbo in memory instead of two.
+    That roughly halves VRAM/RAM and cold-start time, which keeps the page from
+    locking up while the first recording is processed.
+    """
     if _MODELS:
         return _MODELS
 
@@ -62,21 +69,13 @@ def load_models():
 
     processor = WhisperProcessor.from_pretrained(BASE)
 
-    vanilla = WhisperForConditionalGeneration.from_pretrained(
-        BASE, torch_dtype=dtype
-    ).to(device)
-    vanilla.eval()
-
-    tuned_base = WhisperForConditionalGeneration.from_pretrained(
-        BASE, torch_dtype=dtype
-    )
-    tuned = PeftModel.from_pretrained(tuned_base, ADAPTER).to(device)
-    tuned.eval()
+    base = WhisperForConditionalGeneration.from_pretrained(BASE, torch_dtype=dtype)
+    model = PeftModel.from_pretrained(base, ADAPTER).to(device)
+    model.eval()
 
     _MODELS.update(
         processor=processor,
-        vanilla=vanilla,
-        tuned=tuned,
+        model=model,
         device=device,
         dtype=dtype,
     )
@@ -186,17 +185,17 @@ def diff_html(vanilla_text, tuned_text):
 @spaces.GPU(duration=60)
 def transcribe(audio_path):
     if not audio_path:
-        return "", "", '<span class="fw-placeholder-text">Record or upload some speech to begin.</span>'
+        return "", "", '<span class="fw-placeholder-text">Record or upload some speech, then hit Transcribe.</span>'
 
     m = load_models()
+    model, processor = m["model"], m["processor"]
     audio = _load_audio_16k_mono(audio_path)
 
-    vanilla_text = _decode(
-        m["vanilla"], m["processor"], audio, m["device"], m["dtype"]
-    )
-    tuned_text = _decode(
-        m["tuned"], m["processor"], audio, m["device"], m["dtype"]
-    )
+    # Vanilla = the same weights with the LoRA switched off, so we never hold a
+    # second copy of the base model in memory.
+    with model.disable_adapter():
+        vanilla_text = _decode(model, processor, audio, m["device"], m["dtype"])
+    tuned_text = _decode(model, processor, audio, m["device"], m["dtype"])
     return vanilla_text, tuned_text, diff_html(vanilla_text, tuned_text)
 
 
@@ -237,13 +236,13 @@ html, body, gradio-app, .gradio-container, .gradio-container > .main,
 .gradio-container .prose, .gradio-container p, .gradio-container span,
 .gradio-container label { color: var(--fw-ink-dark); }
 
-/* ---- hero ------------------------------------------------------------- */
+/* ---- hero — kept compact so the recorder sits within the first screen -- */
 .fw-hero {
   background: var(--fw-card);
   border: 3px double var(--fw-royal-gold);
   border-radius: 8px;
-  padding: 3rem 2.5rem;
-  margin: 0.5rem 0 2.2rem;
+  padding: 1.8rem 2.5rem 1.6rem;
+  margin: 0.25rem 0 1.4rem;
   box-shadow: 0 4px 20px rgba(158,122,68,0.04);
   text-align: center;
 }
@@ -251,18 +250,18 @@ html, body, gradio-app, .gradio-container, .gradio-container > .main,
   font-family: 'Instrument Serif', serif;
   font-style: normal;
   text-transform: none;
-  font-size: 1.25rem;
+  font-size: 1.15rem;
   font-weight: 600;
   color: var(--fw-royal-gold);
-  margin: 0 0 1rem;
+  margin: 0 0 0.5rem;
 }
 .fw-hero h1 {
   font-family: 'Instrument Serif', Georgia, serif !important;
   font-weight: 400 !important;
-  font-size: 4rem !important;
-  line-height: 1.05 !important;
+  font-size: 3.2rem !important;
+  line-height: 1.04 !important;
   color: var(--fw-ink-dark) !important;
-  margin: 0 0 0.8rem !important;
+  margin: 0 0 0.3rem !important;
   font-style: normal !important;
 }
 .fw-hero h1 .fw-accent {
@@ -271,19 +270,50 @@ html, body, gradio-app, .gradio-container, .gradio-container > .main,
 }
 .fw-lede {
   font-family: 'Instrument Serif', Georgia, serif !important;
-  font-size: 2rem !important;
+  font-size: 1.55rem !important;
   color: var(--fw-ink-muted) !important;
-  margin: 0 0 1.2rem !important;
+  margin: 0 0 0.6rem !important;
 }
 .fw-tagline {
   font-family: 'Newsreader', Georgia, serif !important;
-  font-size: 1.15rem !important;
-  line-height: 1.65 !important;
+  font-size: 1.1rem !important;
+  line-height: 1.55 !important;
   color: var(--fw-ink-dark) !important;
-  max-width: 60rem;
-  margin: 0 auto 2.2rem !important;
+  max-width: 42rem;
+  margin: 0 auto 1.1rem !important;
 }
 .fw-tagline strong { color: var(--fw-royal-gold); font-weight: 600; }
+
+/* ---- WER stat — the visual anchor of the hero ------------------------- */
+.fw-wer {
+  display: flex;
+  align-items: baseline;
+  justify-content: center;
+  gap: 0.5rem;
+  margin: 0.4rem 0 0.2rem;
+}
+.fw-wer-num {
+  font-family: 'Instrument Serif', Georgia, serif;
+  font-size: 3rem;
+  line-height: 1;
+  color: var(--fw-ink-muted);
+}
+.fw-wer-good { color: var(--fw-royal-gold); font-weight: 600; }
+.fw-wer-arrow { font-size: 2rem; color: var(--fw-ink-muted); }
+.fw-wer-unit {
+  font-family: 'Instrument Serif', Georgia, serif;
+  font-size: 1.6rem;
+  color: var(--fw-ink-dark);
+  letter-spacing: 0.04em;
+}
+.fw-wer-note {
+  font-family: 'Newsreader', Georgia, serif !important;
+  font-size: 0.95rem !important;
+  color: var(--fw-ink-muted) !important;
+  margin: 0 auto 1.1rem !important;
+  max-width: 38rem;
+}
+
 .fw-badges { display: flex; flex-wrap: wrap; justify-content: center; gap: 0.6rem; }
 .fw-badge {
   display: inline-block;
@@ -414,6 +444,15 @@ html, body, gradio-app, .gradio-container, .gradio-container > .main,
   font-size: 1.35rem !important;
 }
 
+/* ---- call-to-action cue pointing at the recorder ---------------------- */
+.fw-cta {
+  font-family: 'Instrument Serif', Georgia, serif !important;
+  font-size: 1.25rem !important;
+  font-weight: 600 !important;
+  color: var(--fw-royal-gold) !important;
+  margin: 1.2rem 0 0 !important;
+}
+
 /* ---- scale audio component fonts so they are not diminished ---- */
 .gradio-container .fw-audio,
 .gradio-container .fw-audio * {
@@ -425,6 +464,16 @@ html, body, gradio-app, .gradio-container, .gradio-container > .main,
 .gradio-container .fw-audio button {
   font-size: 1.35rem !important;
 }
+/* Keep the waveform / playback chrome from spilling out of its box (the
+   stray "1x" rate badge overflowing the player). */
+.gradio-container .fw-audio .waveform-container,
+.gradio-container .fw-audio .controls,
+.gradio-container .fw-audio .component-wrapper {
+  max-width: 100% !important;
+  overflow: hidden !important;
+  box-sizing: border-box !important;
+}
+.gradio-container .fw-audio .controls * { line-height: normal !important; }
 
 .fw-placeholder-text {
   color: var(--fw-ink-dark) !important;
@@ -484,21 +533,28 @@ footer { display: none !important; }
 
 HEADER_HTML = """
 <div class="fw-hero">
-  <div class="fw-eyebrow">✻ Apache-2.0 • 100% Offline • Trained on Modal </div>
+  <div class="fw-eyebrow">✻ Apache-2.0 • Runs offline • Trained on Modal</div>
   <h1><span class="fw-accent">fluent</span>Whisper</h1>
   <p class="fw-lede">Speak messy. Read clean.</p>
-  <p class="fw-tagline">Cuts fillers, repeats, and false starts in a <strong>single pass.</strong> <br>
-  On your laptop. Offline.
-  </p>
-  <p class="fw-lede">9.4% → 3.4% WER</p>
-  <p class="fw-tagline">The only open model that does this</p>
+  <p class="fw-tagline">Cuts the fillers, repeats, and false starts in a
+  <strong>single local pass</strong> — and shows you exactly what it removed.</p>
+
+  <div class="fw-wer">
+    <span class="fw-wer-num">9.4%</span>
+    <span class="fw-wer-arrow">→</span>
+    <span class="fw-wer-num fw-wer-good">3.4%</span>
+    <span class="fw-wer-unit">WER</span>
+  </div>
+  <p class="fw-wer-note">vanilla Whisper vs. this adapter on DisfluencySpeech — the only open model that does this</p>
 
   <div class="fw-badges">
     <span class="fw-badge">whisper-large-v3-turbo</span>
     <span class="fw-badge">Disfluency correction</span>
-    <span class="fw-badge">LoRA Finetuned</span>
+    <span class="fw-badge">LoRA finetuned</span>
     <span class="fw-badge gold">Apache-2.0</span>
   </div>
+
+  <p class="fw-cta">Record yourself just below and watch it clean up ↓</p>
 </div>
 """
 
@@ -565,23 +621,32 @@ def build_demo():
             type="filepath",
             label="Speak or upload audio",
             elem_classes=["fw-audio"],
+            show_share_button=False,
+            # The animated live-recording waveform renders continuously on the
+            # browser's main thread and is the prime suspect for the "page
+            # unresponsive" freeze mid-recording. Turn it off; the static
+            # playback waveform still appears once recording stops.
+            waveform_options=gr.WaveformOptions(show_recording_waveform=False),
         )
         run_btn = gr.Button(
             "Transcribe", variant="primary", elem_classes=["fw-btn"]
         )
 
-        with gr.Row():
-            vanilla_out = gr.Textbox(
-                label="Vanilla Whisper", lines=4, elem_classes=["fw-out"]
-            )
-            tuned_out = gr.Textbox(
-                label="Cleaned (LoRA)", lines=4, elem_classes=["fw-out"]
-            )
+        # The diff is the headline result, so it sits right under the button —
+        # no scrolling past empty boxes to see what the model did.
         diff_out = gr.HTML(
             label="What the LoRA removed",
             elem_classes=["fw-diff"],
-            value='<span class="fw-placeholder-text">Record or upload some speech to begin.</span>',
+            value='<span class="fw-placeholder-text">Record or upload some speech, then hit Transcribe.</span>',
         )
+        with gr.Accordion("Show raw transcripts", open=False):
+            with gr.Row():
+                vanilla_out = gr.Textbox(
+                    label="Vanilla Whisper", lines=4, elem_classes=["fw-out"]
+                )
+                tuned_out = gr.Textbox(
+                    label="Cleaned (LoRA)", lines=4, elem_classes=["fw-out"]
+                )
 
         run_btn.click(
             transcribe,