koby commited on
Commit
4a18a04
·
1 Parent(s): bb8ed57

exact duration: force min=max semantic tokens + end fade, slider to 180s

Browse files
Files changed (1) hide show
  1. app.py +23 -4
app.py CHANGED
@@ -68,6 +68,22 @@ def flac_to_wav(flac_path, wav_path):
68
  return wav_path
69
 
70
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
71
  def get_duration_seconds(audio_path):
72
  try:
73
  import soundfile as sf
@@ -134,12 +150,14 @@ def generate(style, lyrics, cot, seed, cfg_scale, duration):
134
  duration = int(duration) if duration else 120
135
  except Exception:
136
  duration = 120
137
- duration = max(30, min(duration, 240))
138
  gen_kwargs = dict(style=style, lyrics=lyrics, cot=cot, seed=seed)
139
  if cfg_scale and float(cfg_scale) != 1.0:
140
  gen_kwargs["cfg_scale"] = float(cfg_scale)
141
- # duration control: cap semantic tokens (~25 tokens/sec of audio)
142
- gen_kwargs["semantic_sampling"] = {"max_tokens": duration * 25}
 
 
143
  song = pipeline(**gen_kwargs)
144
 
145
  song_id = secrets.token_hex(6)
@@ -147,6 +165,7 @@ def generate(style, lyrics, cot, seed, cfg_scale, duration):
147
  mp3_path = os.path.join(OUT, f"song_{song_id}.mp3")
148
  wav_path = os.path.join(OUT, f"{song_id}.wav")
149
  song.save(flac_path)
 
150
  flac_to_mp3(flac_path, mp3_path)
151
  flac_to_wav(flac_path, wav_path)
152
 
@@ -183,7 +202,7 @@ with gr.Blocks(title="YuE2-3B Music") as demo:
183
  seed = gr.Number(label="Seed", value=0, precision=0)
184
  cfg = gr.Number(label="CFG scale (1.0 = default)", value=1.0, precision=2)
185
  with gr.Row():
186
- duration = gr.Slider(30, 240, value=120, step=10, label="Duration (max seconds)", scale=4)
187
  duration_txt = gr.Textbox(value="2:00 min", label="≈ time", scale=1, interactive=False)
188
  btn = gr.Button("Generate song", variant="primary")
189
  with gr.Column():
 
68
  return wav_path
69
 
70
 
71
+ def _fade_out(flac_path, total_seconds, fade_dur=2.0):
72
+ """Gentle fade at the very end so a forced stop never clicks. Keeps exact length."""
73
+ try:
74
+ if total_seconds <= fade_dur + 1:
75
+ return
76
+ tmp = flac_path + ".fade.flac"
77
+ subprocess.run(
78
+ ["ffmpeg", "-y", "-i", flac_path, "-af",
79
+ f"afade=t=out:st={total_seconds - fade_dur:.2f}:d={fade_dur}", tmp],
80
+ capture_output=True, check=True,
81
+ )
82
+ os.replace(tmp, flac_path)
83
+ except Exception as e:
84
+ print(f"[fade] skipped: {e}")
85
+
86
+
87
  def get_duration_seconds(audio_path):
88
  try:
89
  import soundfile as sf
 
150
  duration = int(duration) if duration else 120
151
  except Exception:
152
  duration = 120
153
+ duration = max(30, min(duration, 180))
154
  gen_kwargs = dict(style=style, lyrics=lyrics, cot=cot, seed=seed)
155
  if cfg_scale and float(cfg_scale) != 1.0:
156
  gen_kwargs["cfg_scale"] = float(cfg_scale)
157
+ # exact duration: force the model to emit exactly this many semantic
158
+ # tokens (~25 tokens/sec of audio) — min == max, plus a gentle end fade
159
+ n_tokens = duration * 25
160
+ gen_kwargs["semantic_sampling"] = {"min_tokens": n_tokens, "max_tokens": n_tokens}
161
  song = pipeline(**gen_kwargs)
162
 
163
  song_id = secrets.token_hex(6)
 
165
  mp3_path = os.path.join(OUT, f"song_{song_id}.mp3")
166
  wav_path = os.path.join(OUT, f"{song_id}.wav")
167
  song.save(flac_path)
168
+ _fade_out(flac_path, duration)
169
  flac_to_mp3(flac_path, mp3_path)
170
  flac_to_wav(flac_path, wav_path)
171
 
 
202
  seed = gr.Number(label="Seed", value=0, precision=0)
203
  cfg = gr.Number(label="CFG scale (1.0 = default)", value=1.0, precision=2)
204
  with gr.Row():
205
+ duration = gr.Slider(30, 180, value=120, step=10, label="Duration (seconds, exact)", scale=4)
206
  duration_txt = gr.Textbox(value="2:00 min", label="≈ time", scale=1, interactive=False)
207
  btn = gr.Button("Generate song", variant="primary")
208
  with gr.Column():