YuE2 Official SheetSage2 Transcribe:
SomoraYuE2Transcribe uses the official SheetSage2 transcription models to extract a musical score from a recording. This can help you turn existing music into score-based material for a score-conditioned cover workflow. It analyzes musical content rather than preserving the original waveform or copying a singer’s voice, so it is not a voice-cloning or voice-conversion tool. It requires the separate SheetSage runtime and its official transcription models.
YuE2 Official SheetSage2 Transcribe Input Parameters:
audio
The AUDIO recording to transcribe. Provide one nonempty recording from an upstream audio node. The node writes the recording to a temporary WAV file for transcription and deletes that file afterward. There is no default value; the recording is required.
mode
Chooses between melody and full. melody sends a melody-only request to the transcription runtime, while full requests transcription without that restriction. Choose melody when the melodic line is your focus. The interface does not specify a default.
device
Chooses where transcription runs: cuda:0, cuda:1, or cpu. The CUDA options select GPU devices, while cpu runs without a CUDA GPU. Select a CUDA device only if that device is available in your setup. The interface does not specify a default.
dtype
Selects the numeric precision passed to the transcription runtime: bf16 or fp32. This setting can affect runtime memory use and compatibility, although the provided node source does not establish which choice is best for a particular system or recording. The interface does not specify a default.
preset
Passes either default or paper to the transcription runtime. These are named runtime presets; the provided source does not define their internal settings or establish a quality difference. Use default unless your workflow calls for the paper preset. The interface does not specify a default.
max_seconds
Limits how much of the source recording is submitted for transcription, in seconds. The default is 0, which the node converts to no length limit so the entire source can be transcribed. A positive value sets a limit. The minimum is 0, the maximum is 7200, and the interface changes the value in steps of 1.
overlap_seconds
Sets the overlap duration passed to the runtime when you turn off automatic_window_settings. Overlap concerns how adjacent portions of a long recording share audio, but the provided source does not define the runtime’s windowing algorithm. The default is 200 seconds, the minimum is 0, the maximum is 1000, and the step is 1. With automatic window settings enabled, the node passes no explicit overlap value, so changing this field does not affect the submitted setting.
lookahead_seconds
Sets the lookahead duration passed to the runtime when you turn off automatic_window_settings. It is a window-processing setting; the provided source does not specify how the runtime applies it to the score. The default is 100 seconds, the minimum is 0, the maximum is 1000, and the step is 1. With automatic window settings enabled, the node passes no explicit lookahead value.
automatic_window_settings
Controls whether the runtime receives explicit overlap and lookahead durations. The default is True: the node passes neither manual value, leaving those settings to the runtime. Set it to False when you need the overlap_seconds and lookahead_seconds values you entered to be used.
prompts
An optional string of transcription task names. Leave it blank, as it is by default, for all official tasks. If you specify tasks, separate their names with commas; the node trims spaces and passes the resulting names as a list. The source notes that task names can include timestamp, but it does not provide a complete list of accepted names, so use only names supported by your SheetSage runtime.
export_logits
An optional Boolean setting, False by default, that asks the runtime to export logits when enabled. Logits are intermediate model values rather than the score itself. The provided source does not specify their storage location or format; enable this only when your workflow needs that additional data.
export_scores
An optional Boolean setting, False by default, that asks the runtime to export scores when enabled. This is a separate runtime export option from the node’s abc output. The provided source does not specify the exported files’ format or location.
export_embeddings
An optional Boolean setting, False by default, that asks the runtime to export embeddings when enabled. Embeddings are intermediate model representations, generally useful for specialist analysis rather than ordinary score use. The provided source does not specify their format or location.
output_hidden_states
An optional Boolean setting, False by default, that requests hidden states from the runtime. These are intermediate model representations, not a more detailed version of the visible score. The provided source does not specify how the runtime returns or stores them.
render_audio
An optional Boolean setting, False by default, that requests audio rendering from the transcription runtime. The node still has only its two documented string outputs; the provided source does not establish where any rendered audio is saved or whether it is produced for every task.
render_score
Requests score rendering in one or more formats. The choices are none, pdf, svg, png, and pdf,svg,png; none tells the runtime not to render a score. This setting does not change the two documented output sockets. The interface does not specify a default in its control definition, although the transcription method uses none if no value is supplied.
render_parts
A string passed to the runtime to select parts for rendering. Its default is mix. The provided source does not define other accepted part names or their effects, so retain mix unless you know the part names supported by your runtime.
YuE2 Official SheetSage2 Transcribe Output Parameters:
abc
A STRING containing the ABC score returned by the transcription runtime. ABC is a text-based way to represent musical notation; this output is the score-oriented result to pass to a compatible downstream step or inspect as text. It is not the original audio or a voice recording. The node takes this value from the runtime result’s abc field.
report
A STRING containing an indented JSON report. When the runtime result contains a metadata field, the node formats that metadata; otherwise, it formats the entire result. Use it to inspect details the runtime provides about a particular run, rather than assuming it always contains the same fields. It is informational text, not an additional audio or score-rendering socket.
YuE2 Official SheetSage2 Transcribe Usage Tips:
- Start with
modeset tomelodywhen you need the tune; usefullwhen you want a transcription request that is not melody-only. - Leave
max_secondsat0to process the entire recording. Set a positive value when you only need an initial portion or want to limit the amount of audio submitted. - Keep
automatic_window_settingsenabled unless you have a reason to set overlap and lookahead manually. To apply youroverlap_secondsandlookahead_secondsvalues, turn it off. - Leave
promptsblank for all official tasks. Enter comma-separated task names only when you know which tasks your installed runtime supports. - Keep the optional intermediate-data exports off for an ordinary score workflow. Enable a specific export or rendering option only when you need that result and have a way to inspect what the runtime produces.
- Choose an available
device. If a selected CUDA device is unavailable, try another listed device orcpu; processing on a CPU may take longer.
YuE2 Official SheetSage2 Transcribe Common Errors and Solutions:
Provide one nonempty AUDIO recording
- Explanation: The node rejects audio that is empty or does not have the expected single-recording waveform shape.
- Solution: Connect a valid AUDIO output and check that the upstream step produced a recording with audible or silent samples, rather than an empty result. If you are supplying a batch, pass one recording at a time.
Transcription mode must be melody or full
- Explanation: The node received a mode other than
melodyorfull. - Solution: Select one of the two listed modes. If a saved workflow supplies a different value, update that field and run it again.
Select the official SheetSage2 transcription models
- Explanation: The transcription execution path requires the official
sheetsage2model andmert2model keys. The node normally supplies these keys itself, so this message may indicate a modified or mismatched installation. - Solution: Check that the node package, its models, and its runtime belong to the same supported installation. Restore the official transcription models rather than substituting YuE2 generation models.
Model directory or model files could not be found
- Explanation: The execution path resolves the required model directories before transcription. The exact error wording depends on the path resolver and installation.
- Solution: Install the required official SheetSage2 and MERT2 model assets where the extension expects them, then confirm that your ComfyUI installation can access those directories.
SheetSage runtime could not start or complete transcription
- Explanation: This node calls a separate runtime. A missing dependency, incompatible installation, or runtime failure can stop the run; the exact message comes from that component and is not specified in the node source.
- Solution: Check the full runtime error in your console, confirm that the separate SheetSage runtime is installed, and verify that its version and model assets match your node installation.
CUDA device unavailable or out of memory
- Explanation: Selecting
cuda:0orcuda:1requires that device to be available. A GPU run may also fail if it cannot obtain enough memory; the exact error text depends on the system. - Solution: Select a CUDA device that exists on your machine, close other GPU-heavy tasks, or try
cpu. Limitingmax_secondscan reduce the amount of audio submitted, but the source does not guarantee that it resolves GPU memory errors.
Unsupported dtype, preset, prompt, or rendering option
- Explanation: The node forwards these choices to the separate runtime. The provided source does not define every runtime validation rule, so an incompatible value may produce a runtime-specific error.
- Solution: Return to the node’s listed choices, leave
promptsblank to use all official tasks, and keeprender_partsatmixunless your runtime documents another accepted value.
