FireRedTTS3 Whisper Transcribe

The pack's own README says it bluntly: a correct prompt_text transcript materially improves voice cloning. And here's the kicker - FireRedTTS3VoiceClone ships with prompt_text defaulting to empty. If you run a clone without it, you're leaving quality on the table. This node is the one-click fix: drop your reference clip in, get an exact transcript out, wire it into the clone, done.

It's a utility node and it knows it. It runs an OpenAI Whisper ASR pipeline on whatever AUDIO you feed it and returns two things:

• transcript (STRING) - the text. This is the whole point.
• audio (AUDIO) - the input, passed through unchanged, so you can chain the same clip straight into the clone node without a separate copy.

The wiring move that makes it click: right-click the Voice Clone node, convert prompt_text to an input, and connect this node's transcript there. Then the loop closes - clone in, clean reference out, transcript filled automatically.

The inputs are Whisper controls, and the defaults are sensible:

• model - whisper-large-v3-turbo by default, which is the right call: fast and usually accurate enough for a reference transcript. Drop to whisper-small or tiny if you're tight on memory, or up to full large-v3 if a language is being stubborn. Missing models download into ComfyUI/models/audio_encoders/.
• language - auto detects; setting it explicitly improves accuracy on ambiguous clips. The dropdown covers the usual dozen-plus.
• task - transcribe keeps the clip's language; translate forces English output. Leave it on transcribe for cloning, because the transcript needs to match the reference audio's language for the clone to condition on it properly.
• chunk_length_s - 30 by default for longer clips; 0 lets Transformers decide.
• dtype - auto (bf16 on CUDA/XPU, fp32 otherwise) and download_if_missing true. Both fine as-is.

Two subtleties worth knowing. First, cloning quality depends on the transcript being exact - matching what's actually spoken, including hesitations and filler. Whisper gets close but not perfect, so on a reference clip where precision matters, give the transcript a once-over in a text preview node before trusting it. Second, this node exists because the upstream FireRedTTS3 pipeline uses an LLM API for text normalization, which the pack deliberately excludes - the README is explicit that the pack makes no external API calls. So Whisper here is the fully-local replacement for that, which is a good trade.

One mechanism note buried in the pack source: it stubs out a broken or missing torchcodec at import time so an incompatible wheel can't crash the Transformers audio pipeline. If you see odd torchcodec-related errors elsewhere in your ComfyUI, that's a known rough edge of the audio-in-ComfyUI layer, and this pack at least guards against the worst of it.

Install is the pack-wide story: ComfyUI Manager search "FireRedTTS3-ComfyUI", or git clone https://github.com/Saganaki22/FireRedTTS3-ComfyUI into custom_nodes and run python install.py (needs Transformers 5.3+; install.py never touches torch/transformers itself). The node itself is light - the Whisper download is the only real footprint, and it's a fraction of the multi-gigabyte FireRedTTS3 weights you're already loading. For the single most effective quality upgrade available to a cloning workflow, it's the cheapest node in the pack.