Upload app.py with huggingface_hub

app.py

@@ -12,7 +12,6 @@ from transformers import Wav2Vec2FeatureExtractor, WavLMForXVector
 
 MODEL_ID = "microsoft/wavlm-base-plus-sv"
 TARGET_SR = 16000
-MAX_AUDIO_SEC = 5  # speaker identity stabilizes in 5 sec; shorter = 4x faster on CPU
 
 _feature_extractor = None
 _model = None
@@ -27,14 +26,14 @@ def _load_model():
     return _feature_extractor, _model
 
 
-def _embed(audio_array: np.ndarray, sr: int) -> np.ndarray:
+def _embed(audio_array: np.ndarray, sr: int, max_sec: int) -> np.ndarray:
     fe, mdl = _load_model()
     waveform = torch.tensor(audio_array, dtype=torch.float32)
     if waveform.ndim == 2:
         waveform = waveform.mean(0)
     if sr != TARGET_SR:
         waveform = torchaudio.functional.resample(waveform, sr, TARGET_SR)
-    waveform = waveform[: MAX_AUDIO_SEC * TARGET_SR]
+    waveform = waveform[: max_sec * TARGET_SR]
     inputs = fe(waveform.numpy(), sampling_rate=TARGET_SR, return_tensors="pt")
     with torch.no_grad():
         out = mdl(**inputs)
@@ -44,6 +43,7 @@ def _embed(audio_array: np.ndarray, sr: int) -> np.ndarray:
 def identify_speakers(
     repo_ids_text: str,
     samples_per_book: int,
+    audio_sec: int,
     threshold: float,
     hf_token: str,
     progress=gr.Progress(),
@@ -73,7 +73,7 @@ def identify_speakers(
                 raw = row["audio"]
                 audio_bytes = raw.get("bytes") or open(raw["path"], "rb").read()
                 audio_array, sr = sf.read(io.BytesIO(audio_bytes))
-                embs.append(_embed(audio_array, sr))
+                embs.append(_embed(audio_array, sr, int(audio_sec)))
             if embs:
                 embeddings[repo] = np.mean(embs, axis=0)
             else:
@@ -126,34 +126,69 @@ def identify_speakers(
     return df, summary, "\n".join(errors) if errors else "None"
 
 
-with gr.Blocks(title="Speaker Identifier") as demo:
-    gr.Markdown(
-        """
+DESCRIPTION = """
 # 🎙️ Speaker Identifier
-Extract speaker embeddings from HF audio datasets and cluster into unique speakers.
-Uses **WavLM-Base+** (`microsoft/wavlm-base-plus-sv`) — language-agnostic, works for any language.
+
+Finds unique speakers across multiple HF audio datasets. Each dataset is assumed to have
+**one speaker** (e.g. an audiobook). The app extracts voice embeddings and clusters datasets
+by voice similarity.
+
+**Model:** [microsoft/wavlm-base-plus-sv](https://huggingface.co/microsoft/wavlm-base-plus-sv) —
+language-agnostic speaker embeddings, works for any language.
+
+---
+
+## How to use
+
+1. **Paste dataset repo IDs** (one per line, `owner/name` format) into the left box.
+2. **Adjust parameters** if needed (defaults work well for audiobooks):
+   - *Samples per book* — how many audio chunks to average per dataset. More = more robust, slower.
+   - *Audio length (sec)* — how many seconds of each chunk to use. 5 sec is enough for a clear voice; increase if results look wrong.
+   - *Same-speaker threshold* — cosine similarity cutoff. Raise it if too many books are merged into one speaker; lower it if one person gets split into multiple IDs.
+   - *HF Token* — only needed for **private** repos.
+3. Click **Identify Speakers**.
+
+## Output columns
+
+| Column | Meaning |
+|---|---|
+| `dataset` | Repo name (short) |
+| `speaker_id` | Cluster label — same ID means same voice |
+| `books_with_speaker` | How many books share this speaker |
+| `intra_sim` | Avg cosine similarity between books with the same speaker (1.0 = only one book; lower = cluster is less tight) |
+| `closest_match` | Most similar other book and its similarity score |
+
+**Tip:** Sort by `speaker_id` to see all books by the same narrator grouped together.
 """
-    )
+
+with gr.Blocks(title="Speaker Identifier") as demo:
+    gr.Markdown(DESCRIPTION)
+    gr.Markdown("---")
     with gr.Row():
         with gr.Column(scale=2):
             repo_input = gr.Textbox(
-                label="Dataset repo IDs (one per line, owner/name)",
+                label="Dataset repo IDs (one per line)",
                 placeholder="fosters/some_audiobook_output\nfosters/another_audiobook_output",
-                lines=14,
+                lines=16,
             )
         with gr.Column(scale=1):
             samples = gr.Slider(1, 10, value=3, step=1, label="Samples per book")
+            audio_sec = gr.Slider(
+                2, 30, value=5, step=1,
+                label="Audio length per sample (sec)",
+                info="5 sec is usually enough; longer = more accurate but slower",
+            )
             threshold = gr.Slider(
                 0.60, 0.98, value=0.82, step=0.01,
-                label="Same-speaker threshold (cosine similarity)",
-                info="Higher = stricter matching, more clusters",
+                label="Same-speaker threshold",
+                info="Higher = stricter matching → more clusters",
             )
             hf_token = gr.Textbox(
-                label="HF Token (only for private repos)",
+                label="HF Token (private repos only)",
                 type="password",
                 placeholder="hf_…",
             )
-            run_btn = gr.Button("Identify Speakers", variant="primary")
+            run_btn = gr.Button("Identify Speakers", variant="primary", size="lg")
 
     summary_out = gr.Textbox(label="Summary", interactive=False)
     table_out = gr.Dataframe(
@@ -165,7 +200,7 @@ Uses **WavLM-Base+** (`microsoft/wavlm-base-plus-sv`) — language-agnostic, wor
 
     run_btn.click(
         identify_speakers,
-        inputs=[repo_input, samples, threshold, hf_token],
+        inputs=[repo_input, samples, audio_sec, threshold, hf_token],
         outputs=[table_out, summary_out, errors_out],
     )