FL Qwen3 TTS Transcribe

The single biggest lever on voice-clone quality in this pack isn't a knob on the clone node - it's a transcript you probably don't want to type. Qwen3-TTS's clone in default mode conditions on the reference text, and a transcript that doesn't match the audio quietly degrades the clone. This node exists to fix that: it runs Whisper on your reference clip and hands you the transcription as a string, ready to feed into Voice Clone's ref_text or Voice Clone Prompt's. One wire, and you never have to guess what was said in that 10-second sample again.

It's also just a handy transcription node in general. If you've got any AUDIO in your graph and need its text, this does it locally with OpenAI's Whisper models, no API key involved.

How it works

It loads a Whisper model through transformers - AutoModelForSpeechSeq2Seq plus the processor - so first use downloads the model from HuggingFace and caches it in memory. Your audio gets downmixed to mono, resampled to Whisper's expected 16kHz, run through the model, and the text comes back as a plain STRING.

The default model is openai/whisper-large-v3-turbo, which is a good balance of accuracy and speed. The dropdown also offers large-v3, medium, small, base, and tiny - the smaller ones download fast and use less VRAM at the cost of accuracy, which matters if your reference clip has background noise or an accent Whisper stumbles on. device auto-selects CUDA, then MPS, then CPU, and it uses fp16 where it can.

The inputs that matter

• audio - any ComfyUI AUDIO. Your reference clip.
• model - the Whisper size. Start with the default; drop to whisper-small or base if VRAM is tight.
• language - auto by default, or force one of the ten Qwen3-TTS languages. Forcing the language is a real accuracy win on non-English clips.
• device - auto, cuda, or cpu. Leave on auto.

Output: one transcription (STRING) - wire it straight into Voice Clone's ref_text or Voice Clone Prompt's ref_text.

Installing

Same as the rest of the pack. ComfyUI Manager → "FL Qwen3 TTS", or:

cd ComfyUI/custom_nodes
git clone https://github.com/filliptm/ComfyUI-FL-Qwen3TTS.git
cd ComfyUI-FL-Qwen3TTS
pip install -r requirements.txt

Restart. First transcription pulls the Whisper model (the default large-v3-turbo is roughly 1.6GB) - expect that first run to be slow.

Gotchas

Watch the model size vs. your VRAM. Whisper-large on a card already holding a 1.7B TTS model can spill over; that's exactly when to switch to whisper-small or run the transcribe node on CPU (device = cpu). And proofread: Whisper is good but not perfect, and an error in the transcript becomes an error in the clone. The 15-second truncation that Voice Clone applies to the reference also matters here - transcribe the same portion of audio the clone will actually use. The pack applies that truncation at generation time, so if your clip runs long, keep the transcript to the first 15 seconds.