Upload app.py

app.py

@@ -145,20 +145,128 @@ def generate_voice_design(text, language, voice_description):
     return path
 
 
+def enhance_text_with_emotions(client, text):
+    """Use AI to rewrite text with emotional delivery cues that TTS picks up naturally."""
+    if not client:
+        return text
+    try:
+        response = client.chat.completions.create(
+            model=OMNI_MODEL, modalities=["text"],
+            messages=[
+                {
+                    "role": "system",
+                    "content": (
+                        "You are an audiobook performance coach. Rewrite the text to add natural "
+                        "emotional delivery cues that a TTS voice will pick up from context.\n\n"
+                        "Techniques:\n"
+                        "- Add pause markers: '...' for dramatic pauses\n"
+                        "- Use punctuation for emphasis: '!' for energy, '?' for curiosity\n"
+                        "- Add breath/pacing cues with commas and dashes\n"
+                        "- Extend words for emphasis: 'sooo beautiful'\n"
+                        "- Add interjections: 'Oh!', 'Hmm...', 'Ah,'\n"
+                        "- Use ellipsis for trailing off: 'I never thought...'\n\n"
+                        "Rules:\n"
+                        "1. Keep ALL the original meaning and content\n"
+                        "2. Do NOT add stage directions or [brackets]\n"
+                        "3. The output must be speakable text only\n"
+                        "4. Be subtle — enhance, don't overdo\n"
+                        "5. Output ONLY the enhanced text"
+                    ),
+                },
+                {"role": "user", "content": f"Enhance this for emotional TTS delivery:\n\n{text}"},
+            ],
+        )
+        enhanced = response.choices[0].message.content.strip()
+        print(f"[Emotions] Original: {len(text)} chars -> Enhanced: {len(enhanced)} chars")
+        return enhanced
+    except Exception as e:
+        print(f"[Emotions] Enhancement failed: {e}")
+        return text
+
+
+def extract_pdf_sections(filepath):
+    """Extract sections/chapters from a PDF and return them as a list."""
+    try:
+        import pypdf
+    except ImportError:
+        raise gr.Error("pypdf not installed.")
+
+    reader = pypdf.PdfReader(filepath)
+    full_text = ""
+    for page in reader.pages:
+        t = page.extract_text()
+        if t:
+            full_text += t + "\n\n"
+
+    # Try to detect chapters/sections
+    sections = []
+    # Common patterns: "Chapter X", "CHAPTER X", "Part X", "Section X", numbered headings
+    pattern = r'(?m)^(?:(?:Chapter|CHAPTER|Part|PART|Section|SECTION)\s+\w+[.:)?\s].*|(?:\d+[.)]\s+[A-Z].*)|(?:[IVXLCDM]+[.)]\s+.*))'
+    matches = list(re.finditer(pattern, full_text))
+
+    if matches:
+        for i, match in enumerate(matches):
+            start = match.start()
+            end = matches[i + 1].start() if i + 1 < len(matches) else len(full_text)
+            title = match.group().strip()[:80]
+            content = full_text[start:end].strip()
+            if len(content) > 20:
+                sections.append({"title": title, "content": content, "chars": len(content)})
+    
+    # If no sections found, split by double newlines into chunks
+    if not sections:
+        paragraphs = re.split(r'\n\s*\n', full_text)
+        # Group into ~3000 char sections
+        current_section = ""
+        section_num = 1
+        for para in paragraphs:
+            para = para.strip()
+            if not para:
+                continue
+            if len(current_section) + len(para) > 3000 and current_section:
+                sections.append({
+                    "title": f"Section {section_num}",
+                    "content": current_section.strip(),
+                    "chars": len(current_section),
+                })
+                section_num += 1
+                current_section = para
+            else:
+                current_section += "\n\n" + para
+        if current_section.strip():
+            sections.append({
+                "title": f"Section {section_num}",
+                "content": current_section.strip(),
+                "chars": len(current_section),
+            })
+
+    return sections, full_text
+
+
 @spaces.GPU
-def generate_voice_clone(text, language, ref_audio, ref_text):
+def generate_voice_clone(text, language, ref_audio, ref_text, add_emotions):
     if not text.strip():
         raise gr.Error("Please enter some text.")
     if ref_audio is None:
         raise gr.Error("Please upload a reference audio sample.")
+
+    # Enhance text with emotions if requested
+    final_text = text
+    if add_emotions:
+        client = get_llm_client()
+        if client:
+            final_text = enhance_text_with_emotions(client, text)
+        else:
+            print("[Clone] No LLM client for emotion enhancement, using raw text")
+
     model = get_model("clone")
     lang = language if language != "Auto" else "Auto"
-    kwargs = {"text": text, "language": lang, "ref_audio": ref_audio}
+    kwargs = {"text": final_text, "language": lang, "ref_audio": ref_audio}
     if ref_text and ref_text.strip():
         kwargs["ref_text"] = ref_text.strip()
     else:
         kwargs["x_vector_only_mode"] = True
-    print(f"[TTS] Clone: lang={lang}")
+    print(f"[TTS] Clone: lang={lang}, emotions={'yes' if add_emotions else 'no'}")
     wavs, sr = model.generate_voice_clone(**kwargs)
     path = os.path.join(OUTPUT_DIR, f"clone_{int(time.time())}.wav")
     sf.write(path, wavs[0], sr)
@@ -445,7 +553,16 @@ The old man was silent for a long time. "I already did," he finally whispered. "
 
 DESCRIPTION = """
 # Qwen3-TTS Demo
+### Open-Source Text-to-Speech (1.7B) — Self-Hosted, No API Keys for TTS
+
+| Mode | What it does |
+|------|-------------|
+| **Custom Voice** | Pick a preset voice + optional emotion instruction |
+| **Voice Design** | Describe any voice in natural language |
+| **Voice Clone** | Clone a voice from 3 seconds of audio |
+| **Multi-Speaker Story** | Auto-detect characters, assign voices, add emotions — full audiobook |
 
+10 languages supported. Running on ZeroGPU (free).
 """
 
 with gr.Blocks(title="Qwen3-TTS Demo") as demo:
@@ -497,7 +614,10 @@ with gr.Blocks(title="Qwen3-TTS Demo") as demo:
 
     # ── Tab 3: Voice Clone ──
     with gr.Tab("Voice Clone"):
-        gr.Markdown("Clone any voice from a short audio sample (3+ seconds).")
+        gr.Markdown(
+            "Clone any voice from a short audio sample (3+ seconds).\n"
+            "Enable 'Enhance with emotions' to have AI add natural emotional cues to the text."
+        )
         with gr.Row():
             with gr.Column():
                 vc_text = gr.Textbox(label="Text (in cloned voice)", lines=4, placeholder="What should the cloned voice say...")
@@ -505,11 +625,13 @@ with gr.Blocks(title="Qwen3-TTS Demo") as demo:
                 vc_ref = gr.Audio(label="Reference Audio (3+ seconds)", type="filepath")
                 vc_ref_text = gr.Textbox(label="Transcript (optional, improves quality)",
                                          placeholder="What the person says in the reference...")
+                vc_emotions = gr.Checkbox(value=False, label="Enhance with emotions",
+                                          info="AI adds natural emotional cues (pauses, emphasis, interjections) to make the delivery more expressive. Requires DASHSCOPE_API_KEY.")
                 vc_btn = gr.Button("Clone & Generate", variant="primary")
             with gr.Column():
                 vc_audio = gr.Audio(label="Cloned Voice Speech", type="filepath")
         vc_btn.click(fn=generate_voice_clone,
-                     inputs=[vc_text, vc_lang, vc_ref, vc_ref_text], outputs=vc_audio)
+                     inputs=[vc_text, vc_lang, vc_ref, vc_ref_text, vc_emotions], outputs=vc_audio)
 
     # ── Tab 4: Multi-Speaker Story ──
     with gr.Tab("Multi-Speaker Story"):
@@ -517,8 +639,9 @@ with gr.Blocks(title="Qwen3-TTS Demo") as demo:
             "Paste a story with dialogue and the AI automatically:\n"
             "1. Detects all characters and their genders\n"
             "2. Assigns unique voices (male/female matched)\n"
-            "3. Adds emotion instructions per line (sad, whispered, excited...)\n"
+            "3. Adds emotion instructions per line\n"
             "4. Generates the full audiobook with different voices\n\n"
+            "**PDF users:** Upload a PDF, click 'Detect Sections' to pick a specific chapter.\n\n"
             "*Requires DASHSCOPE_API_KEY for character/emotion analysis. TTS is local.*"
         )
         with gr.Row():
@@ -526,6 +649,18 @@ with gr.Blocks(title="Qwen3-TTS Demo") as demo:
                 ms_text = gr.Textbox(label="Story Text", lines=10, placeholder="Paste your story with dialogue...")
                 ms_file = gr.File(label="Or Upload (.txt, .pdf, .docx)",
                                   file_types=[".txt", ".md", ".pdf", ".docx"], type="filepath")
+
+                # PDF section selection
+                with gr.Accordion("PDF Section Selection (optional)", open=False):
+                    ms_detect_btn = gr.Button("Detect Sections from PDF", variant="secondary", size="sm")
+                    ms_sections_info = gr.Markdown(value="Upload a PDF then click 'Detect Sections' to see chapters.")
+                    ms_section_choice = gr.Dropdown(
+                        choices=[], label="Select Section to Generate",
+                        visible=False, allow_custom_value=False,
+                        info="Pick a chapter/section, or close this accordion to process the full text",
+                    )
+                    ms_load_section = gr.Button("Load Selected Section", variant="secondary", size="sm", visible=False)
+
                 ms_lang = gr.Dropdown(choices=LANGUAGES, value="English", label="Language")
                 ms_sample = gr.Button("Load Sample Story", variant="secondary", size="sm")
                 ms_btn = gr.Button("Generate Multi-Speaker Audiobook", variant="primary", size="lg")
@@ -535,14 +670,57 @@ with gr.Blocks(title="Qwen3-TTS Demo") as demo:
                 with gr.Accordion("Full Transcript (with emotions)", open=False):
                     ms_transcript = gr.Markdown()
 
+        # Section detection state
+        _section_cache = gr.State(value=[])
+
+        def detect_sections_from_pdf(file_input):
+            if file_input is None:
+                raise gr.Error("Upload a PDF first.")
+            ext = os.path.splitext(file_input)[1].lower()
+            if ext != ".pdf":
+                raise gr.Error("Section detection works with PDF files. For other formats, paste text directly.")
+            sections, _ = extract_pdf_sections(file_input)
+            if not sections:
+                return "No sections detected.", gr.update(visible=False), gr.update(visible=False), []
+            choices = [f"{s['title']} ({s['chars']:,} chars)" for s in sections]
+            info = f"**Found {len(sections)} sections:**\n\n"
+            for i, s in enumerate(sections):
+                preview = s["content"][:100].replace("\n", " ")
+                info += f"{i+1}. **{s['title']}** ({s['chars']:,} chars) — {preview}...\n"
+            return (
+                info,
+                gr.update(visible=True, choices=choices, value=choices[0]),
+                gr.update(visible=True),
+                sections,
+            )
+
+        def load_selected_section(choice, sections):
+            if not sections or not choice:
+                return ""
+            idx = next((i for i, s in enumerate(sections)
+                        if f"{s['title']} ({s['chars']:,} chars)" == choice), 0)
+            return sections[idx]["content"]
+
+        ms_detect_btn.click(
+            fn=detect_sections_from_pdf, inputs=[ms_file],
+            outputs=[ms_sections_info, ms_section_choice, ms_load_section, _section_cache],
+        )
+        ms_load_section.click(
+            fn=load_selected_section, inputs=[ms_section_choice, _section_cache],
+            outputs=[ms_text],
+        )
+
         ms_sample.click(fn=lambda: SAMPLE_STORY, outputs=ms_text)
         ms_btn.click(fn=generate_multi_speaker_story,
                      inputs=[ms_text, ms_file, ms_lang],
                      outputs=[ms_audio, ms_stats, ms_transcript])
 
+
     gr.Markdown(
         "---\n"
+        "**Models:** Qwen3-TTS-12Hz-1.7B (Apache 2.0) | "
         "**Languages:** EN, ZH, JA, KO, DE, FR, RU, PT, ES, IT | "
+        "**TTS:** Self-hosted on ZeroGPU | "
         "**Character Analysis:** Cloud AI (optional, for Multi-Speaker mode)"
     )