Load Reference Audio:
Drama3LoadReferenceAudio, displayed as Load Reference Audio in the audio/drama3 category, prepares a recorded voice clip for use as a voice-cloning reference in a Drama 3 workflow. It can remove silence from the beginning and end of a recording and shorten a long recording by cutting at a pause before a chosen duration limit. This helps you create a focused reference that ends between words rather than abruptly interrupting speech, following the preparation approach described for Fish Audio’s Drama 3 Space. You can also load a recording without this trimming. The built-in player lets you hear the original file before execution and the prepared clip after execution, making it easier to check what will be passed to later stages. This node prepares audio only: it does not transcribe speech, create an encoded voice reference, generate new speech, or remove background noise.
Load Reference Audio Input Parameters:
audio
A required COMBO selection identifying the recording to load, with support for uploading audio through the interface. The available options are dynamically populated from files recognized as audio or video in ComfyUI’s input directory and sorted by filename; actual decoding support depends on the audio loader. Select a recording containing the voice you want to use as a reference. The node checks that the selected file exists before execution and then loads its waveform and sample rate. It also fingerprints the file contents so replacing a recording under the same filename can be recognized as a change. There is no explicit default filename and no numerical minimum or maximum. For predictable results, choose a clear recording whose speech is easy to hear and whose file can be played successfully.
trim
A required BOOLEAN control with the options True and False; the default is True. When enabled, the node first removes silent edges and then applies the selected maximum reference duration, using a pause-aware cut for a recording that needs shortening. This creates a more compact reference while aiming to stop between words. When disabled, the node loads the audio without these trimming operations, and length has no effect on the resulting duration. Disabling trimming is useful when you have already edited your reference precisely or want to compare the original recording with the automatically prepared version. This setting does not enable noise reduction or general shortening of pauses throughout the recording.
length
A required COMBO control that sets the maximum clip duration when trim is enabled. The exact options are 5 s, 10 s, 15 s, 20 s, and 30 s; the default is 20 s, the smallest selectable limit is 5 seconds, and the largest is 30 seconds. A recording that exceeds the selected limit is cut at its latest pause before that limit, so the result can be shorter than the selected duration. This is a duration ceiling, not a target: a short recording is not extended to fill it. The supplied guidance states that Fish Audio recommends 10–30 seconds of reference audio and that its Drama 3 Space uses a 20-second limit. Clean, fluent speech matters more than maximizing duration, and a shorter reference also makes a shorter downstream prompt. This setting is ignored when trim is disabled.
Load Reference Audio Output Parameters:
AUDIO
The prepared recording as a standard ComfyUI AUDIO value, suitable for connection to later nodes that accept audio, such as reference transcription or voice-encoding stages. It contains waveform, the audio samples with an added batch dimension, and sample_rate, the sampling rate returned by the audio loader. With trimming enabled, the waveform represents the edge-trimmed and duration-limited clip; with trimming disabled, it represents the loaded recording without those operations. This node does not explicitly resample audio to the Drama 3 codec’s 44.1 kHz rate, and its output is ordinary audio rather than encoded voice data or a transcript. The node also provides an audio preview of this same prepared clip, allowing you to listen before using it downstream.
Load Reference Audio Usage Tips:
- Start with
trimenabled andlengthset to20 s. This matches the default reference-preparation duration described for Fish Audio’s Drama 3 Space. - Choose clean, fluent speech with minimal background noise or overlapping voices. Trimming removes silent edges and limits duration, but it does not clean a noisy recording or isolate one speaker.
- Listen to the preview after execution. Check that the first word remains clear, the ending sounds natural, and enough representative speech remains for your reference.
- Use
10 sor15 swhen a shorter excerpt already captures the voice clearly and you want a shorter reference prompt. Use30 swhen useful speech extends beyond the shorter limits, then check whether the extra material improves your workflow. - Set
trimtoFalsewhen your recording is already edited exactly as you want it. Remember that even a selected5 slimit will not shorten the recording while trimming is off. - Transcribe the prepared AUDIO rather than assuming that a transcript of the full recording still matches. Removing part of a recording can leave the original transcript longer than the speech actually supplied to voice encoding.
- If the automatic endpoint does not suit your recording, manually prepare an excerpt that ends at a natural pause, upload it, and disable trimming to preserve your edit.
Load Reference Audio Common Errors and Solutions:
Invalid audio file: filename
- Explanation: This is the node’s explicit validation error, with
filenamereplaced by the selected file reference. It means ComfyUI cannot find the selected file at its resolved input location. - Solution: Upload the recording again or select an existing file from the
audiodropdown. If you imported a workflow from another computer, supply the corresponding recording locally.
FileNotFoundError or “No such file or directory”
- Explanation: A missing-file error can occur if the recording is moved or deleted after selection or validation. Both audio loading and file-content fingerprinting require access to the actual file; the exact message depends on the failing operation.
- Solution: Restore or re-upload the recording, select it again, and rerun the node. Avoid moving or deleting selected input files while a workflow is running.
PermissionError or “Permission denied”
- Explanation: The node needs to read the selected recording and create the input directory if it does not already exist. An operating-system permission error can prevent these operations.
- Solution: Make sure the account running ComfyUI can read the recording and access the input directory. If necessary, upload a copy into ComfyUI’s normal input location or ask your administrator to correct the folder permissions.
Audio loading or decoding failure
- Explanation: The underlying audio loader may reject an unsupported, damaged, or incorrectly labeled file, or a video file without a usable audio track. The supplied node code does not define a fixed error message for these failures, and appearing in the dropdown does not guarantee that a file can be decoded.
- Solution: Check that the file plays in an audio editor, then export a fresh copy in a commonly supported format such as WAV and upload it again. For video sources, extract a valid audio recording first.
The clip remains longer than the selected length
- Explanation: This is not necessarily an execution error. The
lengthsetting is applied only whentrimis enabled; disabling trimming leaves the loaded recording uncut. - Solution: Enable
trim, choose the desiredlength, and execute the node again. Listen to the post-execution preview rather than relying on the original-file preview.
The prepared clip is shorter than the selected length
- Explanation: The selected duration is a maximum, not an exact output length. Removing silent edges can shorten the recording, and choosing an earlier pause can make the result shorter than the limit.
- Solution: Preview the result and keep it if the speech remains suitable. If you need more material, choose a longer limit or manually prepare a longer excerpt with a natural ending.
The transcript does not match the prepared audio
- Explanation: This node does not generate or update transcripts. A transcript written for the full recording may include words removed during trimming.
- Solution: Transcribe the node’s resulting AUDIO or revise your transcript to match the prepared clip exactly before using it in a voice-reference workflow.
