UI und Textcleaning Anpassung

app.py

@@ -1,6 +1,5 @@
 # Thorsten-Voice CosyVoice3 Space
 # Basiert auf FunAudioLLM/Fun-CosyVoice3-0.5B (Apache 2.0)
-# SFT-Inferenz mit dem Thorsten-Voice Finetune
 
 import spaces
 import os
@@ -8,7 +7,7 @@ import sys
 import gradio as gr
 import numpy as np
 import torch
-import random
+import librosa
 
 ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
 sys.path.append(os.path.join(ROOT_DIR, "third_party", "Matcha-TTS"))
@@ -16,122 +15,194 @@ sys.path.append(os.path.join(ROOT_DIR, "third_party", "Matcha-TTS"))
 from huggingface_hub import snapshot_download, hf_hub_download
 
 # ---------------------------------------------------------------------------
-# Modell-Download beim Start (einmalig, wird gecacht)
+# Modell-Download beim Start
 # ---------------------------------------------------------------------------
 MODEL_DIR = "pretrained_models/CosyVoice3-0.5B"
 
-print("Lade Basis-Modell (FunAudioLLM/Fun-CosyVoice3-0.5B-2512) ...")
-snapshot_download(
-    "FunAudioLLM/Fun-CosyVoice3-0.5B-2512",
-    local_dir=MODEL_DIR,
-)
+print("Lade Basis-Modell ...")
+snapshot_download("FunAudioLLM/Fun-CosyVoice3-0.5B-2512", local_dir=MODEL_DIR)
 
 print("Lade Thorsten-Voice Finetune-Gewichte ...")
 for filename in ["llm.pt", "flow.pt", "spk2info.pt"]:
-    hf_hub_download(
-        repo_id="Thorsten-Voice/CosyVoice3",
-        filename=filename,
-        local_dir=MODEL_DIR,
-    )
+    hf_hub_download(repo_id="Thorsten-Voice/CosyVoice3", filename=filename, local_dir=MODEL_DIR)
 
 from cosyvoice.cli.cosyvoice import AutoModel as CosyVoiceAutoModel
 from cosyvoice.utils.common import set_all_random_seed
+from german_transliterate.core import GermanTransliterate
 
 # ---------------------------------------------------------------------------
 # Konstanten
 # ---------------------------------------------------------------------------
-TARGET_SR = 24000
-MAX_TEXT_LEN = 300
-SPEAKER = "thorsten"  # Name aus spk2info.pt
+TARGET_SR  = 24000
+MAX_TEXT_LEN = 2000
+SPEAKER    = "thorsten"
+FIXED_SEED = 42        # Reproduzierbar, für Nutzer unsichtbar
+
+# Trimm-Parameter gegen Artefakte am Anfang
+TOP_DB     = 40
+HOP_LENGTH = 220
+WIN_LENGTH = 440
+
+transliterator = GermanTransliterate(replace={';': ',', ':': ','}, sep_abbreviation=' -- ')
 
+# ---------------------------------------------------------------------------
+# Hilfsfunktionen
+# ---------------------------------------------------------------------------
 
-def generate_seed():
-    return {"__type__": "update", "value": random.randint(1, 100_000_000)}
+def normalize_text(text: str) -> str:
+    return transliterator.transliterate(text)
 
+def trim_audio(audio: np.ndarray, sr: int) -> np.ndarray:
+    """Entfernt stille/artefakt-reiche Bereiche am Anfang und Ende."""
+    _, indices = librosa.effects.trim(
+        audio, top_db=TOP_DB, frame_length=WIN_LENGTH, hop_length=HOP_LENGTH
+    )
+    return audio[indices[0]:indices[1]]
 
 # ---------------------------------------------------------------------------
 # Inferenz
 # ---------------------------------------------------------------------------
 
 @spaces.GPU
-def generate_audio(tts_text: str, seed: int):
+def generate_audio(tts_text: str):
     default_audio = (TARGET_SR, np.zeros(TARGET_SR, dtype=np.float32))
 
     tts_text = tts_text.strip()
     if not tts_text:
         gr.Warning("Bitte gib einen Text ein.")
-        return default_audio
+        return default_audio, ""
+
     if len(tts_text) > MAX_TEXT_LEN:
-        gr.Warning(f"Der Text ist zu lang. Bitte maximal {MAX_TEXT_LEN} Zeichen eingeben.")
-        return default_audio
+        gr.Warning(f"Der Text ist zu lang (max. {MAX_TEXT_LEN} Zeichen).")
+        return default_audio, ""
+
+    normalized = normalize_text(tts_text)
 
-    set_all_random_seed(int(seed))
-    speech_chunks = []
-    for chunk in cosyvoice.inference_sft(
-        tts_text,
-        SPEAKER,
-        stream=False,
-        speed=1.0,
-    ):
-        speech_chunks.append(chunk["tts_speech"])
+    set_all_random_seed(FIXED_SEED)
+    chunks = []
+    for chunk in cosyvoice.inference_sft(normalized, SPEAKER, stream=False, speed=1.0):
+        chunks.append(chunk["tts_speech"])
 
-    audio = torch.concat(speech_chunks, dim=1).numpy().flatten()
-    return (TARGET_SR, audio)
+    audio = torch.concat(chunks, dim=1).numpy().flatten()
+    audio = trim_audio(audio, TARGET_SR)
 
+    return (TARGET_SR, audio), normalized
+
+# ---------------------------------------------------------------------------
+# CSS (Corporate Design)
+# ---------------------------------------------------------------------------
+
+CSS = """
+:root {
+    --tv-dark:  #515f7f;
+    --tv-light: #91a0bf;
+    --tv-yellow:#ffc038;
+}
+
+/* Hintergrund & Schrift */
+body, .gradio-container {
+    background-color: #f5f7fa !important;
+    font-family: 'Segoe UI', Arial, sans-serif !important;
+}
+
+/* Überschrift */
+.tv-header {
+    background: linear-gradient(135deg, var(--tv-dark) 0%, var(--tv-light) 100%);
+    border-radius: 12px;
+    padding: 24px 28px 18px 28px;
+    margin-bottom: 8px;
+    color: white !important;
+}
+.tv-header h1 { color: white !important; margin: 0 0 6px 0; font-size: 1.6rem; }
+.tv-header p  { color: #dde3ef !important; margin: 0; font-size: 0.92rem; }
+.tv-header a  { color: var(--tv-yellow) !important; text-decoration: none; }
+.tv-header a:hover { text-decoration: underline; }
+
+/* Primär-Button */
+button.primary {
+    background: var(--tv-yellow) !important;
+    color: #1a1a1a !important;
+    border: none !important;
+    font-weight: 700 !important;
+    border-radius: 8px !important;
+}
+button.primary:hover {
+    background: #e6a800 !important;
+}
+
+/* Labels */
+label span { color: var(--tv-dark) !important; font-weight: 600; }
+
+/* Normalisierter Text Box */
+.normalized-box textarea {
+    background: #eef1f7 !important;
+    color: #3a4560 !important;
+    font-size: 0.88rem !important;
+}
+
+/* Audio-Output */
+.audio-out { border-top: 3px solid var(--tv-yellow) !important; border-radius: 8px; }
+"""
 
 # ---------------------------------------------------------------------------
 # Gradio UI
 # ---------------------------------------------------------------------------
 
 def build_ui():
-    with gr.Blocks(title="Thorsten-Voice · CosyVoice3") as demo:
-        gr.Markdown(
-            "# 🎙️ Thorsten-Voice · CosyVoice3\n"
-            "Deutsches Text-to-Speech · Finetune von "
-            "[FunAudioLLM/Fun-CosyVoice3-0.5B](https://huggingface.co/FunAudioLLM/Fun-CosyVoice3-0.5B-2512) "
-            "auf dem [Thorsten-Voice Datensatz](https://huggingface.co/thorsten-voice)  \n"
-            "Modell: [Thorsten-Voice/CosyVoice3](https://huggingface.co/Thorsten-Voice/CosyVoice3)"
-        )
+    with gr.Blocks(css=CSS, title="Thorsten-Voice · CosyVoice3") as demo:
+
+        gr.HTML("""
+        <div class="tv-header">
+            <h1>🎙️ Thorsten-Voice &nbsp;·&nbsp; CosyVoice3</h1>
+            <p>
+                Deutsches Text-to-Speech &nbsp;·&nbsp;
+                Finetune von <a href="https://huggingface.co/FunAudioLLM/Fun-CosyVoice3-0.5B-2512" target="_blank">FunAudioLLM/Fun-CosyVoice3-0.5B</a>
+                auf dem <a href="https://huggingface.co/thorsten-voice" target="_blank">Thorsten-Voice Datensatz</a>
+                &nbsp;·&nbsp;
+                Modell: <a href="https://huggingface.co/Thorsten-Voice/CosyVoice3" target="_blank">Thorsten-Voice/CosyVoice3</a>
+            </p>
+        </div>
+        """)
 
         tts_text = gr.Textbox(
-            label="Text (max. 300 Zeichen)",
+            label=f"Text eingeben (max. {MAX_TEXT_LEN} Zeichen)",
             placeholder="Hallo, ich bin Thorsten. Schön, dass du da bist.",
-            lines=4,
+            lines=5,
             value="Hallo, ich bin Thorsten. Schön, dass du da bist.",
+            max_lines=12,
         )
 
-        with gr.Row():
-            seed_button = gr.Button("🎲", scale=0)
-            seed = gr.Number(value=0, label="Zufalls-Seed", scale=1)
-
         generate_button = gr.Button("🔊 Audio generieren", variant="primary")
+
         audio_output = gr.Audio(
             label="Synthetisiertes Audio",
             autoplay=True,
             streaming=False,
+            elem_classes=["audio-out"],
+        )
+
+        normalized_text = gr.Textbox(
+            label="Normalisierter Text (nach german_transliterate)",
+            interactive=False,
+            lines=3,
+            elem_classes=["normalized-box"],
         )
 
-        seed_button.click(fn=generate_seed, inputs=[], outputs=[seed])
         generate_button.click(
             fn=generate_audio,
-            inputs=[tts_text, seed],
-            outputs=[audio_output],
+            inputs=[tts_text],
+            outputs=[audio_output, normalized_text],
         )
 
     return demo
 
-
 # ---------------------------------------------------------------------------
 # Hauptprogramm
 # ---------------------------------------------------------------------------
 
 if __name__ == "__main__":
     print("Initialisiere CosyVoice3 ...")
-    cosyvoice = CosyVoiceAutoModel(
-        model_dir=MODEL_DIR,
-        load_trt=False,
-        fp16=False,
-    )
+    cosyvoice = CosyVoiceAutoModel(model_dir=MODEL_DIR, load_trt=False, fp16=False)
 
     demo = build_ui()
     demo.queue(default_concurrency_limit=2).launch()

requirements.txt

@@ -1,4 +1,4 @@
---extra-index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/onnxruntime-cuda-12/pypi/simple/ # https://github.com/microsoft/onnxruntime/issues/21684
+--extra-index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/onnxruntime-cuda-12/pypi/simple/
 --extra-index-url https://pypi.nvidia.com
 --trusted-host download.pytorch.org
 --trusted-host aiinfra.pkgs.visualstudio.com
@@ -42,3 +42,4 @@ wget==3.2
 funasr==1.2.0
 onnxconverter-common
 x-transformers==2.3.5
+german-transliterate