webui/app.py

@@ -485,6 +485,65 @@ def load_character(name: str):
     return info, ref_path
 
 
+def list_character_voices(name: str) -> dict[str, str]:
+    """label -> reference audio path, for every voice available for a
+    character: the base voice, every approved emotion_library entry, and
+    every pending (not yet approved) emotion-cache candidate. Used both by
+    the clone-as-character tool below and could back a similar picker
+    anywhere else a character's variants need listing."""
+    cards = read_character_cards()
+    card = cards.get(name)
+    if not card or "_error" in card:
+        return {}
+    voices = {}
+    ref_audio = card.get("reference_audio")
+    if ref_audio:
+        voices["Base (natural voice)"] = str(SERVICE_STATION_DIR / ref_audio)
+    for emo, entry in (card.get("emotion_library") or {}).items():
+        ref = entry.get("reference_audio")
+        if ref:
+            voices[f"{emo} (approved)"] = str(SERVICE_STATION_DIR / ref)
+    cache_dir = card["_dir"] / "emotion-cache"
+    if cache_dir.is_dir():
+        for p in sorted(cache_dir.glob("*.wav")):
+            voices[f"{p.stem} (cached, not approved)"] = str(p)
+    return voices
+
+
+def refresh_voice_choices(name: str):
+    voices = list_character_voices(name)
+    labels = list(voices.keys())
+    return gr.update(choices=labels, value=labels[0] if labels else None)
+
+
+def clone_as_character(character_name: str, voice_label: str, text: str, language: str):
+    if not character_name:
+        raise gr.Error("Pick a character first.")
+    if not voice_label:
+        raise gr.Error("Pick a voice (base or one of its emotions).")
+    if not text or not text.strip():
+        raise gr.Error("Enter some text for them to say.")
+
+    voices = list_character_voices(character_name)
+    reference = voices.get(voice_label)
+    if not reference:
+        raise gr.Error(f"Could not resolve '{voice_label}' for {character_name} - try Refresh.")
+
+    _free_comfyui_memory()
+    env = None
+    if TTS_LD_LIBRARY_PATH:
+        env = os.environ.copy()
+        env["LD_LIBRARY_PATH"] = TTS_LD_LIBRARY_PATH
+    out_path = WORKDIR / f"clone_{int(time.time())}.wav"
+    _run(
+        [str(TTS_PYTHON), str(TTS_GENERATE), "--text", text.strip(),
+         "--reference", reference, "--language", language, "--output", str(out_path)],
+        cwd=TTS_DIR, env=env,
+    )
+    transcript = _whisper_transcribe(str(out_path))
+    return str(out_path), transcript
+
+
 def create_character(
     name: str, role: str, prose: str, language: str,
     current_ref: str, history: list,
@@ -775,6 +834,20 @@ with gr.Blocks(title="Audio-demo pipeline studio") as demo:
             char_info = gr.Markdown()
             char_audio = gr.Audio(label="Reference audio", type="filepath", interactive=False)
 
+            gr.Markdown(
+                "## Clone text as this character\n"
+                "Say anything in the selected character's voice - the base voice, an "
+                "approved emotion, or a pending (not yet approved) candidate from the "
+                "cache. Useful both for \"how would they say this?\" and for testing an "
+                "emotion candidate with real dialogue instead of the fixed test line below."
+            )
+            clone_voice_dd = gr.Dropdown(label="Voice", choices=list(list_character_voices(None).keys()))
+            clone_text_tb = gr.Textbox(label="Text to speak", lines=3, placeholder="Type anything.")
+            clone_language_dd = gr.Dropdown(LANGUAGES, value="English", label="Language")
+            clone_btn = gr.Button("Clone", variant="primary")
+            clone_audio = gr.Audio(label="Cloned speech", type="filepath", interactive=False)
+            clone_transcript_box = gr.Textbox(label="Whisper transcript")
+
             gr.Markdown("## Generate a new emotion candidate")
             with gr.Row():
                 emotion_char_dd = gr.Dropdown(label="Character", choices=character_names())
@@ -793,10 +866,25 @@ with gr.Blocks(title="Audio-demo pipeline studio") as demo:
                 outputs=[char_dd, emotion_char_dd],
             )
             char_dd.change(load_character, inputs=[char_dd], outputs=[char_info, char_audio])
+            char_dd.change(refresh_voice_choices, inputs=[char_dd], outputs=[clone_voice_dd])
+            clone_btn.click(
+                clone_as_character,
+                inputs=[char_dd, clone_voice_dd, clone_text_tb, clone_language_dd],
+                outputs=[clone_audio, clone_transcript_box],
+            )
             gen_emotion_btn.click(
                 generate_emotion_candidate,
                 inputs=[emotion_char_dd, emotion_word_tb, emotion_model_dd],
                 outputs=[emotion_ref_audio, emotion_preview_audio, emotion_status],
+            ).then(
+                # So a freshly-generated candidate is immediately pickable in
+                # "Clone text as this character" without reselecting the
+                # character to force a refresh. Keyed off char_dd (the
+                # browser above), not emotion_char_dd - the clone tool
+                # always uses whichever character char_dd is showing, and
+                # those two dropdowns can legitimately point at different
+                # characters at once.
+                refresh_voice_choices, inputs=[char_dd], outputs=[clone_voice_dd],
             )
             # Dropdown choices are otherwise fixed at server-start time, so a
             # character created earlier in the same running session (or by