FL FishSpeech Transcribe

The Reference Audio node needs the exact transcript of your voice clip, and unless you enjoy typing every reference clip by hand, you want this node first. It's a Whisper transcription wrapper: feed it any AUDIO, and it returns the words as a plain string you can wire straight into Reference Audio. It saves you a workflow step and, more importantly, saves you from the typos that wreck voice clones - because a transcript that doesn't match the audio is the fastest way to make the TTS node sound confused.

How it works

Under the hood it loads an OpenAI Whisper model through the transformers library, resamples your audio down to Whisper's expected 16kHz, and runs speech-to-text. The loaded Whisper model is cached in-process, so repeated calls don't re-download or re-load. The default model is whisper-large-v3-turbo - the modern sweet spot of accuracy and speed - and you can step down through large-v3, medium, small, base, and tiny if you want faster transcription on weaker hardware.

Inputs

• audio (required) - a ComfyUI AUDIO input, same format Reference Audio wants, so the same Load Audio output feeds both.
• model (required) - the six Whisper variants above. Default large-v3-turbo is usually the right call; drop to small or base for long clips when you just need a decent draft.
• language (required) - auto plus ten explicit choices: en, zh, ja, ko, de, fr, es, pt, ru, it. Leave it on auto and it figures it out; pin it if auto-detection keeps guessing wrong on accented or mixed speech.
• device (optional) - auto / cuda / cpu, defaulting to auto. On Apple Silicon, auto also finds MPS.

Output

One output, transcription (a plain STRING). Wire it into Reference Audio's transcript input and you've got a fully automated clone pipeline: load clip → transcribe → encode reference → TTS. Or route it anywhere else that eats a string - a display node, a text file saver, whatever.

Common issues

• transformers is not installed - the pack's requirements.txt doesn't list it, and this node is the one that needs it. The code even raises a RuntimeError telling you to install it. If you hit that:

  pip install transformers

  Most ComfyUI installs already have it, but a fresh environment may not.
• First run downloads Whisper weights - a few hundred MB to a couple GB depending on which model you picked, straight from HuggingFace. It's a one-time cost; the model is then cached.
• Garbled transcription on noisy clips - Whisper is resilient but not a miracle worker. The cleaner the reference audio, the cleaner the transcript, and the better your voice clone. This is the same advice the Reference Audio node gives, because it's the same failure.

Installing

Search "FL FishSpeech" in ComfyUI Manager and install, or clone https://github.com/filliptm/ComfyUI-FL-FishSpeech.git into custom_nodes/ and pip install -r requirements.txt. Then restart. The pack-wide setup you only do once: the fish-speech repo at ComfyUI/fish-speech/ (a sibling of custom_nodes/) and a huggingface-cli login so the gated OpenAudio model can download on first use.