ROCm Seed Voice Conversion

Voice conversion is not text-to-speech, and conflating the two is why people bounce off this node. TTS invents a performance from a script; this keeps the one you already have - words, timing, breaths, and melody if it's a sung take - and swaps the timbre underneath. Feed it a recording and a voice, get the recording back in that voice.

Where it sits in the audio pile

Audio is the thinnest layer in ComfyUI: node packs bolted onto a stack that was never built for sound. TTS is the crowded, essentially solved corner (Chatterbox, Kokoro, F5). Voice conversion is a different job with a thinner toolchain. The long-standing open answer is RVC, which needs a trained model per voice. Seed-VC is zero-shot - a few seconds of reference, no training run - so if you want to audition nine voices in an afternoon, this is the shortcut.

A voice is also a legally protected thing in a growing pile of jurisdictions - Tennessee's ELVIS Act was the first state law to extend a right of publicity to AI simulations of a person's voice, and several states cover cloned voices explicitly. Get consent.

How it actually works

Two pipelines sit inside the node. Speech mode (f0_condition off) conditions a flow-matching DiT on whisper-small features plus a CAM++ speaker embedding, and BigVGAN vocodes the result at 22.05 kHz. Singing mode (f0_condition on) switches to a 44.1 kHz F0-conditioned model with RMVPE pitch extraction, so the melody survives the voice swap. Long sources are chunked into 30-second windows with a ~745 ms crossfade, which is why prompt_length is a seam knob as much as a speed knob.

The inputs that matter

Two are required: source_audio (what you keep, downmixed to mono) and reference_audio (the voice to adopt - 3 to 25 seconds, truncated past 25). Then, roughly in order of impact:

• f0_condition - the switch that matters. Off for speech, on for singing; it also decides the output sample rate.
• precision - leave it on fp16. The vocoder dominates runtime, and half precision is the AMD win: ~6x faster vocoding, ~2x end to end on gfx1151.
• steps (30) and inference_cfg_rate (0.7). Under ~16 steps, detail smears. A CFG rate of 0 disables guidance and halves sampling cost.
• prompt_length (10 s) - how much reference goes in as the in-context prompt. Shorter is faster and seams less; timbre and pitch matching still use the full reference.
• separate_source - split vocal from instrumental first (mel-band roformer karaoke) and convert only the vocal.

Singing-only: auto_f0_adjust (matches source pitch level to the reference, on), pitch_shift (±24 semitones, for when the reference sits much lower or higher), vibrato_transfer (1.0 copies the source's vibrato and glides; 0 for the raw model). speed rescales output duration, and unload_model frees the models after the run - leave it off, reloading costs ~30 seconds.

Outputs and wiring

audio is the converted take, same length as the source - wire it to SaveAudio. accompaniment is the instrumental when separate_source is on; the core AudioMerge node puts the song back together. With separation off it's a silent same-length track rather than None, so leaving it wired breaks nothing.

LoadAudio(source) ──> source_audio
LoadAudio(ref)    ──> reference_audio   ──> ROCmSeedVoiceConversion ──> SaveAudio / AudioMerge

Installing it

ComfyUI Manager, search rocm-ninodes - or clone it. It's built for AMD/ROCm but isn't AMD-only; it uses whatever torch device it finds.

cd ComfyUI/custom_nodes
git clone https://github.com/iGavroche/rocm-ninodes.git ComfyUI-ROCM-Optimized-VAE
cd ComfyUI-ROCM-Optimized-VAE && python install.py

install.py pip-installs the pack's requirements (numpy, Pillow, psutil, gguf, safetensors, plus librosa and scipy for this node); torch, torchaudio, transformers and PyYAML come from your ComfyUI environment. separate_source additionally wants pip install audio-separator, deliberately absent from requirements. Coming from v1, run uv run python tools/upgrade_to_v2.py first, then restart completely. The node lands under 🎛 ROCm/audio.

The models are the real install

Seven things have to exist under ComfyUI/models/TTS/: four Plachta/Seed-VC checkpoints (DiT_seed_v2_uvit_whisper_small_wavenet_bigvgan_pruned.pth, DiT_seed_v2_uvit_whisper_base_f0_44k_bigvgan_pruned_ft_ema_v2.pth, campplus_cn_common.bin, rmvpe.pt), the openai/whisper-small folder, and the two nvidia BigVGAN vocoder folders (bigvgan_v2_22khz_80band_256x, bigvgan_v2_44khz_128band_512x). hf download <repo> --local-dir models/TTS/… fetches each.

Two traps. The BigVGAN folders must physically exist - a missing path is treated as a Hugging Face repo id and fails. And the singing checkpoint must be the _ft_ema_v2 file; the old name isn't loaded any more. With separate_source on, the ~1.7 GB karaoke checkpoint downloads itself into models/TTS/Seed-VC/separator/ on first use.

Where it goes wrong

The node isn't in the menu. The console prints ROCmSeedVoiceConversion unavailable: … when a vendored import fails - usually transformers, torchaudio or yaml missing from the ComfyUI environment. One bad import hides the node instead of exploding it, so read the console before reinstalling anything.

FileNotFoundError under models/TTS/Seed-VC. One of the four checkpoints isn't where it expects. By far the most common failure, and always a path problem.

Garbage pitch on a full mix in singing mode. RMVPE will happily track the bass line and the converted vocal wanders. Turn on separate_source, or split upstream, so only the vocal reaches the pitch tracker.

precision=fp16 doing nothing. It's conditional on a GPU device; on CPU or MPS the vocoder stays fp32 and you're in for a long wait. This is a GPU node pretending to be polite about it.