Select podcast voices on the Create page

Add a "Cast" picker on the Create step with one voice dropdown per speaker
(Nova/Atlas/Echo/Sage/Vivi/Onyx/Rune), shown per the speaker-count slider.
The picked voices drive the actual TTS: each script speaker maps to that
voice's reference clip (tts/voices/samples/<id>.wav), so episodes use the
chosen voices and hosts sound distinct. Selected names are also passed as
the script's speaker labels when distinct, so dialogue reads "Nova: …".

Removes the old step-2 per-speaker preset/clone assignment (voices are now
chosen up front); setup_voices just parses lines + speakers.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

app.py

@@ -17,7 +17,6 @@ from typing import List
 import gradio as gr
 
 from research.graph import generate_script, parse_script
-from tts import voices as voice_lib
 from tts import engine as tts_engine
 
 MAX_SPEAKERS = 4
@@ -75,6 +74,26 @@ def _sample_path(voice_id):
     return p if os.path.isfile(p) else None
 
 
+# Selectable library voices (those with a bundled sample), by display name.
+_VOICE_NAME_TO_ID = {name: vid for vid, name, *_ in VOICE_LIBRARY if vid}
+LIBRARY_VOICE_NAMES = list(_VOICE_NAME_TO_ID.keys())
+_VOICE_DEFAULTS = ["Nova", "Atlas", "Echo", "Sage"]
+
+
+def _voice_config_for(name):
+    """Resolve a library voice name (e.g. 'Nova') to its cloning reference clip + text."""
+    vid = _VOICE_NAME_TO_ID.get(name)
+    if not vid:
+        return tts_engine.VoiceConfig()
+    wav = os.path.join(_SAMPLES_DIR, f"{vid}.wav")
+    txt = os.path.join(_SAMPLES_DIR, f"{vid}.txt")
+    ref_text = ""
+    if os.path.isfile(txt):
+        with open(txt, encoding="utf-8") as f:
+            ref_text = f.read().strip()
+    return tts_engine.VoiceConfig(ref_audio=wav if os.path.isfile(wav) else None, ref_text=ref_text)
+
+
 # --------------------------------------------------------------------- HTML builders
 def _avatar(letter: str, idx: int, size: int = 30) -> str:
     grad = _AVATAR_GRADIENTS[idx % len(_AVATAR_GRADIENTS)]
@@ -202,48 +221,41 @@ def _episode_meta(topic: str, script: str, bed: str) -> str:
 
 
 # ------------------------------------------------------------------- Phase 1 handler
-def run_research(topic, style, duration, num_speakers, progress=gr.Progress()):
+def run_research(topic, style, duration, num_speakers, *voice_names, progress=gr.Progress()):
     topic = (topic or "").strip()
     if not topic:
         raise gr.Error("Please enter a topic for the podcast.")
+    n = int(num_speakers)
+    # Use the picked voice names as the script's speaker names when they're distinct, so the
+    # dialogue reads "Nova: …" / "Atlas: …". (Audio mapping is order-based regardless, so a
+    # duplicate pick simply falls back to generic Host/Guest labels.)
+    names = [voice_names[i] for i in range(n) if i < len(voice_names) and voice_names[i]]
+    speaker_names = names if (len(names) == n and len(set(names)) == n) else None
     progress(0.1, desc="Planning research…")
     state = generate_script(
         topic,
         style=style,
         duration_min=int(duration),
-        num_speakers=int(num_speakers),
+        num_speakers=n,
+        speaker_names=speaker_names,
     )
     progress(0.9, desc="Finalizing script…")
     script = state.get("script", "")
     return script, script, script_to_bubbles(script)
 
 
-# ------------------------------------------------------- set up per-speaker voice rows
+# ------------------------------------------------- parse script into lines + speakers
 def setup_voices(script_text):
     lines = parse_script(script_text or "")
     speakers: List[str] = []
     for spk, _ in lines:
         if spk not in speakers:
             speakers.append(spk)
-    speakers = speakers[:MAX_SPEAKERS]
-
-    updates = [lines, speakers]  # lines_state, speakers_state
-    presets = voice_lib.list_presets()
-    for i in range(MAX_SPEAKERS):
-        if i < len(speakers):
-            updates.append(gr.update(visible=True))
-            updates.append(gr.update(value=f"### 🎤 Voice for “{speakers[i]}”"))
-            default_choice = presets[min(i + 1, len(presets) - 1)] if len(presets) > 1 else presets[0]
-            updates.append(gr.update(choices=presets, value=default_choice))
-        else:
-            updates.append(gr.update(visible=False))
-            updates.append(gr.update())
-            updates.append(gr.update())
-    return updates
+    return lines, speakers[:MAX_SPEAKERS]
 
 
 # ------------------------------------------------------------------- Phase 2 handler
-def run_tts(lines, speakers, topic, bed, *voice_inputs, progress=gr.Progress()):
+def run_tts(lines, speakers, topic, bed, *voice_names, progress=gr.Progress()):
     if not tts_engine.is_available():
         raise gr.Error(
             "The TTS stack (fish-speech + GPU) is not available in this environment. "
@@ -253,20 +265,15 @@ def run_tts(lines, speakers, topic, bed, *voice_inputs, progress=gr.Progress()):
         raise gr.Error("No script is loaded. Generate a script, then review it before "
                        "producing the episode.")
 
-    presets = list(voice_inputs[0:MAX_SPEAKERS])
-    clones = list(voice_inputs[MAX_SPEAKERS:2 * MAX_SPEAKERS])
-    texts = list(voice_inputs[2 * MAX_SPEAKERS:3 * MAX_SPEAKERS])
-
+    # Map each speaker to the voice picked on the Create page. Prefer a name match (the
+    # script's speaker labels are the picked voice names), else fall back to slot order.
     voice_map = {}
     for i, speaker in enumerate(speakers):
-        clone_path = clones[i]
-        if clone_path:  # cloning overrides preset
-            voice_map[speaker] = tts_engine.VoiceConfig(
-                ref_audio=clone_path, ref_text=(texts[i] or "").strip()
-            )
+        if speaker in _VOICE_NAME_TO_ID:
+            voice_map[speaker] = _voice_config_for(speaker)
         else:
-            v = voice_lib.get_voice(presets[i])
-            voice_map[speaker] = tts_engine.VoiceConfig(ref_audio=v.ref_audio, ref_text=v.ref_text)
+            name = voice_names[i] if i < len(voice_names) else (voice_names[0] if voice_names else None)
+            voice_map[speaker] = _voice_config_for(name)
 
     progress(0.05, desc="Analyzing script & pacing…")
     sr, audio = tts_engine.generate_podcast(lines, voice_map, progress=progress)
@@ -589,6 +596,25 @@ def build_ui():
                                     [b[0] for b in SOUND_BEDS], value="Ambient Drift", label="SOUND BED",
                                 )
 
+                        # ---- Cast: pick the voice used for each speaker ----
+                        with gr.Column(elem_classes=["pf-card"]):
+                            gr.HTML(
+                                "<div class='pf-step-h'><span class='t'>🎙 Cast</span>"
+                                "<span class='h'>pick the voice for each speaker — "
+                                "preview them on the Voices page</span></div>"
+                            )
+                            with gr.Row():
+                                voice_pickers = []
+                                for i in range(MAX_SPEAKERS):
+                                    default = (_VOICE_DEFAULTS[i] if i < len(_VOICE_DEFAULTS)
+                                               else LIBRARY_VOICE_NAMES[i % len(LIBRARY_VOICE_NAMES)])
+                                    vp = gr.Dropdown(
+                                        LIBRARY_VOICE_NAMES, value=default,
+                                        label=f"Speaker {i + 1}", visible=(i < 2),
+                                        elem_classes=["pf-config-card"],
+                                    )
+                                    voice_pickers.append(vp)
+
                         with gr.Row():
                             gr.HTML("<div style='flex:1'></div>")
                             generate_script_btn = gr.Button("⌁  Generate script", elem_classes=["pf-cta"], scale=0)
@@ -599,7 +625,7 @@ def build_ui():
                             gr.HTML(
                                 "<div class='pf-step-h'><span class='num'>2</span>"
                                 "<span class='t'>Review the script</span>"
-                                "<span class='h'>Edit any line · tap a speaker to reassign</span></div>"
+                                "<span class='h'>Edit any line · voices are set on the Create step</span></div>"
                             )
                             regenerate_btn = gr.Button("↻  Regenerate", elem_classes=["pf-ghost"], scale=0)
 
@@ -617,33 +643,6 @@ def build_ui():
                                 show_label=False, elem_id="pf-bed", container=False,
                             )
 
-                        gr.HTML("<div class='pf-section-label'>VOICE ASSIGNMENT</div>")
-                        preset_dropdowns, clone_audios, clone_texts = [], [], []
-                        speaker_rows, speaker_labels = [], []
-                        for i in range(MAX_SPEAKERS):
-                            with gr.Group(visible=False, elem_classes=["pf-card"]) as row:
-                                speaker_label = gr.Markdown(f"### Speaker {i + 1}")
-                                with gr.Row():
-                                    preset = gr.Dropdown(
-                                        voice_lib.list_presets(),
-                                        value=voice_lib.list_presets()[0],
-                                        label="Preset voice",
-                                    )
-                                    clone = gr.Audio(
-                                        sources=["upload", "microphone"],
-                                        type="filepath",
-                                        label="…or clone a voice (10–30s reference)",
-                                    )
-                                ref_text = gr.Textbox(
-                                    label="Reference transcript (optional, improves cloning)",
-                                    lines=1,
-                                )
-                            speaker_rows.append(row)
-                            speaker_labels.append(speaker_label)
-                            preset_dropdowns.append(preset)
-                            clone_audios.append(clone)
-                            clone_texts.append(ref_text)
-
                         with gr.Row():
                             back_btn = gr.Button("←  Back to topic", elem_classes=["pf-ghost"], scale=0)
                             gr.HTML("<div style='flex:1'></div>")
@@ -729,10 +728,13 @@ def build_ui():
         # keep the bubble preview in sync with raw-script edits
         script_box2.change(script_to_bubbles, inputs=[script_box2], outputs=[review_bubbles])
 
-        setup_outputs = [lines_state, speakers_state]
-        for i in range(MAX_SPEAKERS):
-            setup_outputs += [speaker_rows[i], speaker_labels[i], preset_dropdowns[i]]
+        # show one voice picker per selected speaker
+        num_speakers.change(
+            lambda n: [gr.update(visible=i < int(n)) for i in range(MAX_SPEAKERS)],
+            inputs=[num_speakers], outputs=voice_pickers,
+        )
 
+        setup_outputs = [lines_state, speakers_state]
         step_views = [step1, step2, step3, step_loading]
 
         # Step 1 -> loading -> generate -> Step 2
@@ -742,7 +744,7 @@ def build_ui():
             outputs=step_views + [loading_panel],
         ).then(
             run_research,
-            inputs=[topic, style, duration, num_speakers],
+            inputs=[topic, style, duration, num_speakers] + voice_pickers,
             outputs=[script_box, script_box2, review_bubbles],
         ).then(
             setup_voices, inputs=[script_box2], outputs=setup_outputs,
@@ -759,7 +761,7 @@ def build_ui():
             outputs=step_views + [loading_panel],
         ).then(
             run_research,
-            inputs=[topic, style, duration, num_speakers],
+            inputs=[topic, style, duration, num_speakers] + voice_pickers,
             outputs=[script_box, script_box2, review_bubbles],
         ).then(
             setup_voices, inputs=[script_box2], outputs=setup_outputs,
@@ -774,8 +776,7 @@ def build_ui():
             outputs=step_views + [loading_panel],
         ).then(
             run_tts,
-            inputs=[lines_state, speakers_state, topic, bed]
-            + preset_dropdowns + clone_audios + clone_texts,
+            inputs=[lines_state, speakers_state, topic, bed] + voice_pickers,
             outputs=[audio_out, file_out, result_meta, result_transcript],
         ).then(lambda: show_step(3), outputs=step_views)