Load Qwen3-ASR:
Qwen3ASRLoader, displayed as “Load Qwen3-ASR” in the audio/qwen3_asr category, prepares a Qwen3-ASR speech-recognition model for use in your ComfyUI workflow. Speech recognition converts spoken audio into written text, which is useful when you need a transcript of a voice reference for a Drama 3 audio workflow. The node supports the 1.7B and 0.6B Qwen3-ASR variants in the format produced by this extension’s conversion tool. It loads the model’s learned weights and configuration, prepares the tokenizer that handles text, and integrates the model with ComfyUI’s device and memory management. Its purpose is model preparation rather than transcription: loading the model does not listen to audio or produce a transcript by itself. This separates model setup from audio processing so you can use the prepared model in a compatible transcription stage.
Load Qwen3-ASR Input Parameters:
model_name
A required COMBO selection that identifies the converted Qwen3-ASR model file you want to load. Its options are populated dynamically from ComfyUI’s registered asr model folder, described by the node as models/asr; the source does not define a fixed list of filenames. Select a compatible converted 0.6B or 1.7B checkpoint, depending on which files you have installed. Your selection determines the model configuration and weights that will be available for subsequent speech recognition, rather than directly controlling transcription language or decoding settings. The file must include the qwen3_asr_config metadata created by the extension’s tools/convert.py asr conversion process; an unconverted checkpoint is not sufficient. Options: the files returned by folder_paths.get_filename_list("asr"). Default: no explicit default is declared in the schema. Minimum and maximum: not applicable because this is a file-selection dropdown, not a numeric parameter.
Load Qwen3-ASR Output Parameters:
QWEN3_ASR_MODEL
A custom model object of type QWEN3_ASR_MODEL that makes the loaded Qwen3-ASR recognizer available to compatible downstream nodes. It contains the speech-recognition model, its tokenizer, and the ComfyUI model-management wrapper used to handle execution and device placement. Connect it to a downstream input that accepts this exact model type when you want to transcribe audio. It is not an audio clip, transcript, or standard image-generation model, so it cannot be used interchangeably with those data types. A successful output means model loading and tokenizer initialization have completed; it does not indicate that any audio has been processed.
Load Qwen3-ASR Usage Tips:
-
Convert your Qwen3-ASR checkpoint using this extension’s
tools/convert.py asrworkflow before loading it. Simply copying or renaming an original checkpoint does not add the configuration metadata this loader requires. -
Store the converted file in
ComfyUI/models/asror another location registered for theasrmodel category. If your newly added file is not listed, refresh ComfyUI’s model list or restart ComfyUI. -
For a memory-constrained setup, consider the supported 0.6B variant first. Its smaller parameter count generally reduces model storage and memory requirements compared with the 1.7B variant; the supplied code does not establish a specific accuracy or speed difference between them.
-
Connect the model output to a compatible Qwen3-ASR transcription node and provide your audio there. This loader has no audio, language, timestamp, or decoding controls, so those aspects cannot be configured here.
-
Allow ComfyUI to manage device placement and precision. The loader automatically selects its processing device, sets up an offload device, and chooses bfloat16 computation when ComfyUI considers it appropriate, otherwise float32. There is no manual precision or device selector on this node.
-
If you are planning longer transcription jobs, the accompanying ASR implementation describes processing audio as 16 kHz mono and splitting recordings longer than 20 minutes at quiet points. These are downstream transcription behaviors, not settings or duration limits of the loader itself.
Load Qwen3-ASR Common Errors and Solutions:
“[filename] is not a converted Qwen3-ASR file (tools/convert.py asr)”
- Explanation: The file has no metadata or is missing the required
qwen3_asr_configentry. This usually means you selected an original, unconverted checkpoint or a file intended for a different model. - Solution: Convert the appropriate Qwen3-ASR checkpoint with this extension’s
tools/convert.py asrworkflow, place the result in the registeredasrfolder, and select that converted file.
“Qwen3-ASR decoders with another head size or an untied LM head are not supported”
- Explanation: The checkpoint’s decoder configuration does not match the architecture supported by this implementation. The code checks the attention head size and, for a normal ASR model, requires shared text input and output embeddings.
- Solution: Use a supported Qwen3-ASR 0.6B or 1.7B checkpoint converted with the extension’s tools. Avoid manually modifying the configuration to bypass this check, because incompatible weights will not become compatible through a configuration edit.
“Qwen3-ASR's added tokens did not land on their ids”
- Explanation: The tokenizer’s added tokens were not assigned the identifiers expected by the implementation. These identifiers must match the model so that audio and text markers are interpreted correctly.
- Solution: Update ComfyUI and the extension to compatible versions, and restore any modified tokenizer assets or the extension’s
qwen3_asr/added_tokens.jsonfile. If files are missing or damaged, reinstall the affected component rather than editing token identifiers manually.
Model file not found or unavailable in model_name
- Explanation: The dropdown only lists files discovered in the registered
asrmodel locations. Execution also resolves the selected filename again, so a file moved or deleted after selection cannot be loaded. The exact missing-file message depends on ComfyUI’s folder-path handling. - Solution: Confirm that the converted checkpoint exists in
ComfyUI/models/asror your registered ASR model directory. Refresh or restart ComfyUI and select the filename again.
Configuration parsing errors, such as “JSONDecodeError” or “KeyError”
- Explanation: The loader reads
qwen3_asr_configas JSON and expects configuration fields such astext_configandaudio_config. Invalid JSON or missing fields can prevent model construction, even if the metadata entry exists. - Solution: Recreate the converted checkpoint from an intact source model using the extension’s conversion tool. Keep the converter and loader from the same extension version where possible, and avoid editing checkpoint metadata by hand.
“Error(s) in loading state_dict” or “size mismatch”
- Explanation: Model weights are loaded with strict compatibility checks. Missing weights, unexpected weights, or tensor dimensions that do not match the configuration can cause loading to fail.
- Solution: Verify that the checkpoint is a supported Qwen3-ASR model and that its weights and configuration came from the same source model. Reconvert it with the extension’s tools; if you suspect an incomplete download, obtain a fresh source checkpoint first.
Checkpoint reading errors from PyTorch or safetensors
- Explanation: The checkpoint reader may fail before model validation if a file is incomplete, damaged, inaccessible, or not a supported checkpoint format. Exact messages depend on the file and the underlying reader.
- Solution: Check that the file finished downloading or converting and that ComfyUI has permission to read it. Replace damaged files and use the actual converted checkpoint rather than an unrelated file with a changed extension.
“CUDA out of memory” or another memory-allocation error
- Explanation: Available device or system memory may be insufficient while reading or preparing the checkpoint. The larger 1.7B variant generally requires more memory than the 0.6B variant, and other loaded models can reduce the memory available.
- Solution: Try the supported 0.6B model, close other GPU-heavy applications, and unload unnecessary models or restart ComfyUI. Check both GPU memory and system RAM if the problem persists; this node does not provide a manual memory or precision setting.
