Qwen3-ASR Transcribe:
Qwen3ASRTranscribe converts spoken or sung audio into written text using Qwen3-ASR speech recognition. Its node description advertises support for 30 languages and 22 Chinese dialects, making it useful for preparing dialogue transcripts, documenting vocal performances, and supplying transcription text for Drama 3 voice-encoding workflows. It can automatically detect the language or recognize speech in a language you specify, with optional contextual guidance to help it recognize names and specialized vocabulary. Internally, it prepares audio as 16 kHz mono, converts the sound into features the model can analyze, and generates a transcription using greedy decoding, which selects the most likely next token rather than randomly sampling alternatives. Recordings longer than 20 minutes are split at quiet points before recognition. This node focuses on transcription rather than translation or word-level timing.
Qwen3-ASR Transcribe Input Parameters:
model
Required connection of type QWEN3_ASR_MODEL containing the loaded speech-recognition model. Connect it from the Load Qwen3-ASR node, which supports converted Qwen3-ASR 0.6B or 1.7B models stored in the models/asr directory. The model supplies the learned recognition capabilities used to interpret your recording, so a compatible model must be loaded before transcription can run. Model files must be converted with the extension’s tools/convert.py asr workflow; an arbitrary language model or an unconverted checkpoint is not interchangeable. A Qwen3-ForcedAligner model is also unsuitable here because it assigns timing to existing text rather than transcribing speech. There is no default, numeric range, or model-selection list on this parameter; model selection happens in the loader.
audio
Required connection of type AUDIO containing the recording to transcribe, including its waveform and sample rate. The transcription pipeline prepares the recording as 16 kHz mono, so you do not need to supply audio already in that format. Recordings longer than 1,200 seconds, or 20 minutes, are split at quiet points for processing. If the audio contains multiple batch items, each item is transcribed separately and its text is joined with the others using newline separators; audio channels are not treated as separate batch items. Clear, intelligible vocals give the model a better basis for recognition, while overlapping speech or prominent background sound can make the words harder to distinguish. This parameter has no default value or exposed numeric range.
language
Required COMBO setting that controls whether the model detects the language or uses a specified language for recognition. The default is auto, which allows language detection and attempts to return the detected language alongside the transcript. Selecting a named language bypasses detection and uses that language as the recognition setting; the returned language then reflects your selection rather than an independent detection result. Available options are auto, Chinese, English, Cantonese, Arabic, German, French, Spanish, Portuguese, Indonesian, Italian, Korean, Russian, Thai, Vietnamese, Japanese, Turkish, Hindi, Malay, Dutch, Swedish, Danish, Finnish, Polish, Czech, Filipino, Persian, Greek, Romanian, Hungarian, and Macedonian. Although the node description mentions support for 22 Chinese dialects, those dialects are not individually exposed as additional options in this menu. Choose the known language when appropriate, or use auto when you are unsure; this setting is not a translation-language selector.
context
Required STRING field with multiline support and a default of an empty string, "". Its content is optional guidance for recognition, such as character names, unusual place names, technical terms, or the subject of the recording. For example, Character names: Mira and Kaelen. Topic: restoring an ancient observatory. can provide relevant vocabulary without requiring a complete transcript. This text becomes part of the model’s prompt and may influence how it interprets unclear words, but it does not guarantee exact spelling or replace the audio with your supplied wording. Leave it empty for a baseline transcription, or provide concise, accurate information related to the recording. No minimum length, maximum length, or predefined options are specified in the node interface.
max_new_tokens
Required advanced INT setting that limits how many new text tokens the model may generate for each processed audio segment. The default is 4096, the minimum is 1, and the maximum is 32768. Tokens are units of text that may represent words, parts of words, or punctuation, so this value is not a direct word count or duration limit. The interface describes the allowance as applying per 20 minutes of audio; more precisely, it is applied separately to each segment produced by the long-audio splitting process, including shorter recordings processed as one segment. A limit that is too small can cut off the transcript before all speech has been covered. A larger limit provides more room for lengthy transcripts but can allow longer generation and greater resource use; generation can still stop before reaching the limit. This setting does not control language selection, audio quality, or timestamp generation.
Qwen3-ASR Transcribe Output Parameters:
text
STRING containing the recognized speech or singing as plain text. The implementation removes special decoding tokens, separates recognized text from the model’s language metadata when applicable, and applies repetition cleanup. For a long recording, the text from successive processing segments is concatenated directly, without an extra separator inserted between segments. For multiple audio batch items, the resulting transcripts are joined with newline characters. You can inspect this text, use it as a starting point for captions or dialogue notes, or connect it to Drama 3 Encode Voice’s transcript input. It does not contain word-level timestamps, speaker labels, confidence scores, or a structured subtitle file, and you should review recognition accuracy before reusing it in a production workflow.
language
STRING containing the detected language name when language is set to auto, or the supplied language name when you explicitly select one and recognition produces a nonempty result. If detection does not provide a language, this output can be an empty string. Multiple language names can be returned as a comma-separated string when processing segments or batch items produces different language results. Within a recording, consecutive repeated language names are suppressed, but a language can appear again after a different one; across batch items, identical complete language-result strings are deduplicated. Treat this output as a language summary rather than a time-aligned description of language changes, and remember that a manually selected language is not independently verified by this output.
Qwen3-ASR Transcribe Usage Tips:
- Start with
language=auto, an emptycontext, andmax_new_tokens=4096to establish a baseline. Change one setting at a time so you can judge whether it improves recognition. - If you know the recording’s language, select it explicitly to bypass language detection. Keep
autowhen the language is uncertain, and do not use the language menu to request translation. - Add a short, relevant vocabulary guide in
contextfor fictional character names, uncommon locations, or specialist terminology. Avoid unrelated instructions or a large block of text that does not describe the recording. - If a transcript stops before the recording ends, increase
max_new_tokensabove4096as needed, up to32768. Check the result again rather than assuming a larger token allowance automatically improves recognition accuracy. - For short tests, you can lower
max_new_tokens, but leave enough room for all spoken words. A very small value can produce only a fragment even when the audio is clear. - Use clear audio where possible. For difficult music or overlapping dialogue, try a cleaner vocal track or shorter excerpts to make recognition easier to assess.
- Let the node split recordings longer than 20 minutes automatically, but review the text around segment boundaries because segment transcripts are joined without added separators. Manually dividing a recording can make review and correction easier.
- Before feeding the result into Drama 3 Encode Voice’s transcript input, correct missed words and spelling errors so the text accurately matches the performance. Use a separate alignment workflow if you need word timing.
- For batched audio, remember that transcript lines and the language summary are aggregated differently. Process clips individually if you need an unambiguous transcript-and-language pair for every clip.
Qwen3-ASR Transcribe Common Errors and Solutions:
Qwen3-ForcedAligner times a known text (Drama 3 Editor's aligner input); transcribe with Qwen3-ASR
- Explanation: This exact error appears when the connected model is a Qwen3-ForcedAligner rather than a speech-recognition model. The aligner needs existing text and cannot perform transcription through this node.
- Solution: Load a converted Qwen3-ASR 0.6B or 1.7B model with Load Qwen3-ASR and connect that model instead. Keep the ForcedAligner for a separate timing workflow.
[filename] is not a converted Qwen3-ASR file (tools/convert.py asr)
- Explanation: This loader error means the selected file lacks the required
qwen3_asr_configmetadata. It commonly indicates an unconverted checkpoint or an unrelated model file. - Solution: Convert a supported model with the extension’s
tools/convert.py asrworkflow, place the converted file inmodels/asr, and select it in Load Qwen3-ASR.
Qwen3-ASR decoders with another head size or an untied LM head are not supported
- Explanation: The model’s internal configuration does not match the decoder architecture supported by this implementation. This is a model-compatibility issue, not a problem with your audio or recognition settings.
- Solution: Use a supported Qwen3-ASR 0.6B or 1.7B checkpoint converted for this extension. Do not try to resolve the error by changing
language,context, ormax_new_tokens.
Qwen3-ASR's added tokens did not land on their ids
- Explanation: The tokenizer’s extra tokens do not have the identifiers expected by this implementation. The model relies on these identifiers to distinguish audio-related markers and other special tokens.
- Solution: Check that the extension’s
added_tokens.jsonfile is present and that its files have not been mixed across different versions. Update or reinstall the extension consistently and check that your ComfyUI installation includes its expected Qwen2.5 tokenizer assets.
Error(s) in loading state_dict for Qwen3ASRModel
- Explanation: This PyTorch error heading can appear during loading when the saved model weights do not match the configuration. Details such as missing keys, unexpected keys, or size mismatches identify the incompatible parts.
- Solution: Verify that you selected the intended converted model. Reconvert it from a complete supported checkpoint, and avoid combining configuration or weight files from different models.
CUDA out of memory
- Explanation: This runtime error can occur when there is insufficient GPU memory for the model and audio processing. Actual wording varies by device and PyTorch version; long segments and large generation allowances can increase resource demands.
- Solution: Free GPU memory by unloading other models or applications, try the supported 0.6B model if you are using 1.7B, and test a shorter audio excerpt. Reduce an unnecessarily large
max_new_tokensvalue while keeping enough room for the transcript.
Missing required input: model or audio
- Explanation: This describes a workflow-validation problem rather than an exact error string from the supplied implementation. The node requires both a loaded recognition model and a valid ComfyUI audio connection.
- Solution: Connect Load Qwen3-ASR to
modeland an audio-producing node toaudio. Confirm that the upstream nodes load successfully before running transcription.
KeyError: 'sample_rate' or KeyError: 'waveform'
- Explanation: These errors can occur if a custom upstream node supplies an invalid audio object. The implementation expects the audio data to contain both
sample_rateandwaveform. - Solution: Use an upstream node that produces the standard ComfyUI
AUDIOformat. If you are using a custom audio-processing node, verify that it preserves both required fields and the expected waveform structure.
Transcript is incomplete, empty, or inaccurate without an error message
- Explanation: This is a recognition-result problem rather than a specific exception. An incomplete transcript can result from a generation limit that is too low, while silence, unclear vocals, an unsuitable language selection, or misleading context can contribute to empty or inaccurate results. An empty language output alone does not necessarily mean transcription failed.
- Solution: Listen to the recording, test a short clear excerpt, and compare
autowith the correct explicit language. Remove irrelevant context, add only useful vocabulary, and raisemax_new_tokensif the text appears cut off. Review the transcript before using it downstream.
