fix BPM, whisper force cpu and librosa upgrade

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "UltraSinger"
-version = "0.0.13.dev14"
+version = "0.0.13.dev15"
 description = "A tool to create UltraStar karaoke files from audio files"
 requires-python = ">=3.12"
 dependencies = [
@@ -16,8 +16,7 @@ dependencies = [
     "langcodes",
     "language_data",
     "packaging",
-    "librosa>=0.10.2",
-    "numba>=0.59.0",
+    "librosa",
     "swift-f0",
     "pydub",
     "demucs",

src/Settings.py

@@ -11,7 +11,7 @@
 @dataclass
 class Settings:
 
-    APP_VERSION = "0.0.13.dev14"
+    APP_VERSION = "0.0.13.dev15"
     CONFIDENCE_THRESHOLD = 0.6
     CONFIDENCE_PROMPT_TIMEOUT = 4
 

src/UltraSinger.py

@@ -171,6 +171,10 @@ def run() -> tuple[str, Score, Score]:
     # Create process audio
     process_data.process_data_paths.processing_audio_path = CreateProcessAudio(process_data)
 
+    # Get BPM from wav file
+    if not settings.input_file_is_ultrastar_txt:
+        process_data.media_info.bpm = get_bpm_from_file(process_data.process_data_paths.processing_audio_path)
+
     # Detect key
     detected_key, detected_mode = detect_key_from_audio(process_data.process_data_paths.processing_audio_path)
     if process_data.media_info.music_key is None:
@@ -611,15 +615,17 @@ def transcribe_audio(cache_folder_path: str, processing_audio_path: str) -> Tran
     transcription_result = None
     whisper_align_model_string = None
     if settings.transcriber == "whisper":
-        if not settings.whisper_align_model is None: whisper_align_model_string = settings.whisper_align_model.replace("/", "_")
-        transcription_config = f"{settings.transcriber}_{settings.whisper_model.value}_{settings.pytorch_device}_{whisper_align_model_string}_{settings.whisper_batch_size}_{settings.whisper_compute_type}_{settings.language}"
+        if not settings.whisper_align_model is None:
+            whisper_align_model_string = settings.whisper_align_model.replace("/", "_")
+        whisper_device = "cpu" if settings.force_whisper_cpu else settings.pytorch_device
+        transcription_config = f"{settings.transcriber}_{settings.whisper_model.value}_{whisper_device}_{whisper_align_model_string}_{settings.whisper_batch_size}_{settings.whisper_compute_type}_{settings.language}"
         transcription_path = os.path.join(cache_folder_path, f"{transcription_config}.json")
         cached_transcription_available = check_file_exists(transcription_path)
         if settings.skip_cache_transcription or not cached_transcription_available:
             transcription_result = transcribe_with_whisper(
                 processing_audio_path,
                 settings.whisper_model,
-                settings.pytorch_device,
+                whisper_device,
                 settings.whisper_align_model,
                 settings.whisper_batch_size,
                 settings.whisper_compute_type,
@@ -684,7 +690,6 @@ def infos_from_audio_video_input_file() -> tuple[str, str, str, MediaInfo]:
     if song_info.cover_image_data is not None:
         save_image(song_info.cover_image_data, basename_without_ext, song_folder_output_path)
 
-    real_bpm = get_bpm_from_file(ultrastar_audio_input_path)
     return (
         basename_without_ext,
         song_folder_output_path,
@@ -694,7 +699,6 @@ def infos_from_audio_video_input_file() -> tuple[str, str, str, MediaInfo]:
             title=song_info.title,
             year=song_info.year,
             genre=song_info.genres,
-            bpm=real_bpm,
             cover_url=song_info.cover_url,
             audio_extension=audio_ext,
             video_extension=video_ext

src/modules/Audio/bpm.py

@@ -1,18 +1,25 @@
 import librosa
+import soundfile as sf
 
 from modules.console_colors import ULTRASINGER_HEAD, blue_highlighted
 
 
 def get_bpm_from_data(data, sampling_rate):
     """Get real bpm from audio data"""
     onset_env = librosa.onset.onset_strength(y=data, sr=sampling_rate)
-    wav_tempo = librosa.beat.tempo(onset_envelope=onset_env, sr=sampling_rate)
+    wav_tempo = librosa.feature.tempo(onset_envelope=onset_env, sr=sampling_rate)
 
     print(f"{ULTRASINGER_HEAD} BPM is {blue_highlighted(str(round(wav_tempo[0], 2)))}")
     return wav_tempo[0]
 
 
 def get_bpm_from_file(wav_file: str) -> float:
     """Get real bpm from audio file"""
-    data, sampling_rate = librosa.load(wav_file, sr=None)
+    data, sampling_rate = sf.read(wav_file, dtype='float32')
+    # Transpose if stereo to match librosa's expected format
+    if len(data.shape) > 1:
+        data = data.T
+    # Convert to mono if stereo
+    if data.ndim > 1:
+        data = librosa.to_mono(data)
     return get_bpm_from_data(data, sampling_rate)

src/modules/Audio/youtube.py

@@ -116,7 +116,6 @@ def download_from_youtube(input_url: str, output_folder_path: str, cookiefile: s
             input_url, basename_without_ext, song_output, cookiefile
         )
 
-    real_bpm = get_bpm_from_file(audio_file_path)
     return (
         basename_without_ext,
         song_output,
@@ -126,7 +125,6 @@ def download_from_youtube(input_url: str, output_folder_path: str, cookiefile: s
             title=song_info.title,
             year=song_info.year,
             genre=song_info.genres,
-            bpm=real_bpm,
             cover_url=cover_url,
             video_url=input_url,
             audio_extension=audio_ext,

src/modules/ProcessData.py

@@ -19,7 +19,7 @@ class MediaInfo:
     """Media Info"""
     title: str
     artist: str
-    bpm: float
+    bpm: Optional[float] = None
     year: Optional[str] = None
     genre: Optional[str] = None
     language: Optional[str] = None