fix: Play produces no sound — switch to streaming audio + system deps

Root cause: Audio component with streaming=False replaces audio on each
yield, causing rapid restarts with no audible output on fast GPUs.

Fixes:
- Audio component: streaming=True (yields APPEND to playback buffer)
- Ensure float32 dtype on all audio yields
- Add packages.txt with sox/ffmpeg for Qwen TTS system deps
- Add from __future__ import annotations for Python 3.9 compatibility
- Copy bundled audio_cache to runtime dir on startup (HF bucket)

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

app.py

@@ -4,6 +4,7 @@ import math
 import re
 import struct
 import wave
+import numpy as np
 import gradio as gr
 import time
 from pathlib import Path
@@ -504,7 +505,7 @@ with gr.Blocks(title="MomsVoice", css=css_code) as demo:
                         visible=False, interactive=False,
                         label="Audio Stream",
                         type="numpy",
-                        streaming=False,
+                        streaming=True,
                         autoplay=True,
                     )
 
@@ -956,7 +957,7 @@ with gr.Blocks(title="MomsVoice", css=css_code) as demo:
             para_idx = min(int(i * n_paras / total), n_paras - 1) if total > 0 else 0
             is_last = (i + 1 >= total)
             para_status = f"<div style='text-align:center;font-size:10px;color:#4ade80;font-family:monospace;margin-top:8px;'>▶ Paragraph {para_idx + 1} / {n_paras}</div>"
-            yield ((sample_rate, wav), _status_playing, para_status,
+            yield ((sample_rate, wav.astype(np.float32) if hasattr(wav, 'astype') else wav), _status_playing, para_status,
                    gr.Button(visible=False), gr.Button(visible=True), gr.Button(visible=True),
                    gr.Slider(value=para_idx), render_story_text(paras, para_idx),
                    gr.HTML(visible=is_last),

inference_lfm.py

@@ -4,6 +4,8 @@ LFM2.5-Audio-1.5B inference module — end-to-end audio Q&A.
 Replaces the 3-model pipeline (Whisper ASR + Qwen Q&A + Qwen TTS) with a single
 multimodal model that accepts audio/text input and produces audio+text output.
 """
+from __future__ import annotations
+
 import logging
 import threading
 import torch

packages.txt

@@ -0,0 +1,3 @@
+sox
+libsox-fmt-all
+ffmpeg

tts.py

@@ -410,13 +410,13 @@ def generate_audio_stream(
         item = chunk_q.get()
         if item is _SENTINEL:
             if audio_buffer:
-                combined = np.concatenate(audio_buffer)
+                combined = np.concatenate(audio_buffer).astype(np.float32)
                 yield sample_rate, combined, last_idx, n, None
             break
         i, wav, err = item
         if err:
             if audio_buffer:
-                combined = np.concatenate(audio_buffer)
+                combined = np.concatenate(audio_buffer).astype(np.float32)
                 yield sample_rate, combined, last_idx, n, None
                 audio_buffer = []
                 buffer_samples = 0
@@ -442,7 +442,7 @@ def generate_audio_stream(
 
         # Yield when buffer reaches target size or this is the first chunk (fast start)
         if buffer_samples >= _TARGET_SAMPLES or (i == 0 and buffer_samples > 0):
-            combined = np.concatenate(audio_buffer)
+            combined = np.concatenate(audio_buffer).astype(np.float32)
             yield sample_rate, combined, i, n, None
             audio_buffer = []
             buffer_samples = 0

voice_clone.py

@@ -23,6 +23,8 @@ Usage:
     profile_id = create_voice_profile(ref_audio_path, voice_name="Mom")
     wav, sr = synthesize_cloned(text, profile_id)
 """
+from __future__ import annotations
+
 import json
 import logging
 import os