Deploy Kiku-Engine

app.py

@@ -681,7 +681,7 @@ def _words(vocal_path):
     mono, _ = librosa.load(vocal_path, sr=16000, mono=True)
     with torch.inference_mode():
         heard = whisper()({"raw": mono, "sampling_rate": 16000}, return_timestamps=True,
-                          chunk_length_s=30, batch_size=8)
+                          chunk_length_s=30, batch_size=8, generate_kwargs={"task": "transcribe"})
     lines = []
     for chunk in heard.get("chunks") or []:
         text = " ".join((chunk.get("text") or "").split())
@@ -730,9 +730,9 @@ def _read_take(audio, words=False, separate="auto"):
     return found
 
 
-def render_time(abc, style, lyrics, cot, seed, clock=None, duration=0.0, strip_voice=False,
+def render_time(abc, style, lyrics, cot, seed, clock=None, span=0.0, strip_voice=False,
                 split_parts=False, **kwargs):
-    seconds = min(MAX_SCORE, float(duration or 0) or 60)
+    seconds = min(MAX_SCORE, float(span or 0) or 60)
     total = 45 + int(seconds * 0.9) + cold("yue2", 180)
     if clock:
         total += 30 + int(seconds * 0.45) + cold("sheetsage", 90)
@@ -742,8 +742,9 @@ def render_time(abc, style, lyrics, cot, seed, clock=None, duration=0.0, strip_v
 
 
 @spaces.GPU(duration=render_time)
-def _render_score(abc, style, lyrics, cot, seed, clock=None, duration=0.0, strip_voice=False,
+def _render_score(abc, style, lyrics, cot, seed, clock=None, span=0.0, strip_voice=False,
                   split_parts=False):
+    # `span` is the output length. It cannot be called `duration`: ZeroGPU reserves that keyword.
     """YuE2 plays the score; the result is then locked to `clock` when given."""
     import librosa
     job = job_dir()
@@ -764,13 +765,13 @@ def _render_score(abc, style, lyrics, cot, seed, clock=None, duration=0.0, strip
         score = scores.read(abc)
         heard = sync.bar_starts(beats, len(score.bars), len(mono) / 24000.0)
         if heard:
-            warped, lock = sync.lock(audio, tools.SR, score, clock, heard, float(duration))
+            warped, lock = sync.lock(audio, tools.SR, score, clock, heard, float(span))
             report.update(lock)
             if warped is not None:
                 audio, report["locked"] = warped, True
         if not report["locked"]:
             # No trustworthy bar match: start it on the first bar and let it run.
-            audio = tools.shift(audio, float(clock[0]), tools.SR, length=int(float(duration) * tools.SR))
+            audio = tools.shift(audio, float(clock[0]), tools.SR, length=int(float(span) * tools.SR))
     path = tools.write(audio, str(job / "played.wav"))
     parts = [None, None, None]
     if split_parts:
@@ -849,7 +850,7 @@ def restage(audio, style, lyrics, voice, harmony, abc, clock, edits, seed, progr
             words = scores.lyric_tags(score)
             prompt = "Instrumental, %s, %d BPM, %s." % (described, score.bpm, ", ".join(INSTRUMENTAL))
         progress(0.4, desc="YuE2 is playing it")
-        path, _, report = _render_score(text, prompt, words, cot, seed, clock=clock, duration=duration)
+        path, _, report = _render_score(text, prompt, words, cot, seed, clock=clock, span=duration)
         elapsed = round(time.monotonic() - started)
         status = "Finished in %dm %ds / %s / %d BPM" % (elapsed // 60, elapsed % 60,
                                                           scores.key_words(score.bars[0].key), score.bpm)
@@ -882,7 +883,7 @@ def band_score(audio, style, separate, balance, split_parts, abc, clock, edits,
         prompt = "Instrumental, %s, %d BPM, %s." % (described, band.bpm, ", ".join(INSTRUMENTAL))
         progress(0.4, desc="YuE2 is playing your chords")
         path, parts, report = _render_score(text, prompt, scores.lyric_tags(band), cot, seed, clock=clock,
-                                            duration=duration, strip_voice=True,
+                                            span=duration, strip_voice=True,
                                             split_parts=bool(split_parts))
         job = job_dir()
         vocal_path = found["vocal"] if found else audio