Extensions/ComfyUI-Reverb-TTS
ComfyUI Extension

ComfyUI-Reverb-TTS

Local Reverb-TTS speech synthesis, voice cloning, reusable voice profiles, and WAV export nodes for ComfyUI

By Wolfricbr·Created about 17 hours ago·Updated about 15 hours ago· 0
Wolfricbr/ComfyUI-Reverb-TTS
Nodes—
On cloudLocal install
Stars0
Updatedabout 15 hours ago
Readme

ComfyUI-Reverb-TTS

Local speech synthesis, voice cloning, reusable voice profiles, and WAV export for ComfyUI.

🇧🇷 Português · Original Reverb-TTS model · Workflows

Experimental integration · v0.3.9. This is a community ComfyUI integration, not an official Software Mansion release. Models are not downloaded automatically. Install the weights separately and read the license section before using generated audio.

Features

  • Local loading of Reverb-TTS FP32, FP16, INT8, and INT4 weight variants.
  • Zero-shot voice cloning from a short audio clip and its exact transcript.
  • Official preset voices using locally installed JSON files.
  • Save/load reusable custom voice profiles in .npz format.
  • Select saved profiles from a voice-library dropdown.
  • Text-to-speech generation with language, seed, speed and steps controls.
  • Lossless WAV export (PCM 24-bit, PCM 16-bit, or float), with an audio player in the ComfyUI node.

Important: .npz voice profiles are specific to the Reverb-TTS Voice serialization. They are not interchangeable with Fish Audio Drama 3 .safetensors voice files or other TTS models.

Requirements

  • ComfyUI with Python compatible with the official reverb-tts==1.0.0 package.
  • A supported PyTorch environment; CUDA GPU recommended.
  • Manually downloaded Reverb-TTS weights and Higgs Audio V2 tokenizer/codec weights.
  • A reference recording and matching transcript for cloning.

Dependency warning: reverb-tts can change shared Python dependencies. In one Windows ComfyUI environment, installing it introduced a huggingface-hub 2.x versus diffusers 0.37.1 version conflict. Check planned package changes before installing. An import succeeding does not guarantee every other custom node will remain compatible.

Installation

Clone into ComfyUI/custom_nodes:

cd /d C:\AI\ComfyUI\ComfyUI\custom_nodes
git clone https://github.com/Wolfricbr/ComfyUI-Reverb-TTS.git

For a Windows portable ComfyUI installation, preview the dependency changes first:

C:\AI\ComfyUI\python_embeded\python.exe -m pip install --dry-run -r C:\AI\ComfyUI\ComfyUI\custom_nodes\ComfyUI-Reverb-TTS\requirements.txt

If compatible with your environment, install and verify:

C:\AI\ComfyUI\python_embeded\python.exe -m pip install -r C:\AI\ComfyUI\ComfyUI\custom_nodes\ComfyUI-Reverb-TTS\requirements.txt
C:\AI\ComfyUI\python_embeded\python.exe -c "from reverb_tts import TTS, Voice; print('Reverb-TTS import OK')"

These paths are examples for a portable Windows installation. Use your actual ComfyUI Python executable if installed elsewhere. Restart ComfyUI after installing.

Download the models manually

The node does not download weights automatically. By default, this version prefers the Windows path:

C:\AI\ComfyUI-Data\models\reverb_tts\
├── Reverb-TTS\
│   ├── model\
│   │   ├── tokenizer.json
│   │   ├── tokenizer_config.json
│   │   ├── chat_template.jinja
│   │   └── safetensors\
│   │       ├── fp32\
│   │       │   ├── config.json
│   │       │   └── model.safetensors
│   │       ├── fp16\          (optional)
│   │       ├── int8\          (optional)
│   │       └── int4\          (optional)
│   └── voices\               (optional for voice cloning)
│       └── pt_male_standard.json
├── higgs-audio-v2-tokenizer\
│   ├── config.json
│   ├── model.safetensors
│   └── preprocessor_config.json
└── voices_custom\            (created automatically for custom .npz voices)

The loader searches the existing Windows directory, ComfyUI/models/reverb_tts, and the directory specified by the REVERB_TTS_MODELS_DIR environment variable. Select the model from its dropdown. The root must contain both Reverb-TTS/ and higgs-audio-v2-tokenizer/. This project does not use diffusion_models.

Download these components:

  1. Reverb-TTS FP32 weights: config.json and model.safetensors.
  2. Reverb tokenizer files: tokenizer.json, tokenizer_config.json, chat_template.jinja.
  3. Higgs Audio V2 tokenizer/codec: config.json, model.safetensors, preprocessor_config.json.
  4. Optional preset voice JSONs.
  5. Optional FP16, INT8 or INT4 weights from their respective model/safetensors/ folders, with each variant's config.json.

For example, to download the FP32 model, common tokenizer files, and preset voices to your local folders using Hugging Face Hub:

C:\AI\ComfyUI\python_embeded\python.exe -c "from huggingface_hub import snapshot_download; snapshot_download(repo_id='software-mansion/Reverb-TTS', local_dir=r'C:\AI\ComfyUI-Data\models\reverb_tts\Reverb-TTS', allow_patterns=['model/tokenizer.json', 'model/tokenizer_config.json', 'model/chat_template.jinja', 'model/safetensors/fp32/*', 'voices/*.json'])"

The separate codec must also be downloaded:

C:\AI\ComfyUI\python_embeded\python.exe -c "from huggingface_hub import snapshot_download; snapshot_download(repo_id='eustlb/higgs-audio-v2-tokenizer', local_dir=r'C:\AI\ComfyUI-Data\models\reverb_tts\higgs-audio-v2-tokenizer', allow_patterns=['config.json', 'model.safetensors', 'preprocessor_config.json'])"

These commands are one-time manual setup steps, not automatic runtime downloads.

Available nodes

| Node | Function | |---|---| | Reverb · Load Model (Local) | Load locally installed FP32/FP16/INT8/INT4 weights and select CPU/CUDA. | | Reverb · Clone Voice from Audio | Build a voice from an audio reference and matching transcript. | | Reverb · Preset Voice | Select a locally installed official voice preset by name. | | Reverb · Save Voice (.npz) | Save a reusable custom voice profile. | | Reverb · Voice Library (NPZ Selector) | Pick a saved voice from a dropdown. | | Reverb · Generate Speech | Generate AUDIO from text and a voice. | | Reverb · Save WAV | Save a WAV and display an audio preview/player after execution. |

Generation settings

  • language: pt, en, es, fr, de, it, pl
  • speed: 0.5–2.0, step 0.05
  • steps: default 8 (separate from speed)
  • seed: default 42
  • variant: fp32, fp16, int8, int4
  • force_fp32_compute: changes compute dtype; does not change the weight variant

Example workflows

Open a JSON from workflows/ in ComfyUI.

| Workflow | What it does | |---|---| | 01 — Clone and Generate | Reference audio + transcript → cloned speech WAV. | | 02 — Preset Voice | Official local preset + text → WAV. | | 03 — Create and Save Voice | Reference audio + transcript → reusable .npz voice. | | 04 — Load Voice and Generate | Existing custom voice → WAV. | | 05 — Clone, Save and Generate | One execution produces both a voice profile and speech WAV. |

Suggested first test: Workflow 02 with pt_male_standard and a short sentence. Then try Workflow 03 followed by Workflow 04. To refresh the saved voice list, reload the ComfyUI page or recreate the selector node.

Voice library

Custom profiles are stored in:

C:\AI\ComfyUI-Data\models\reverb_tts\voices_custom\

Use Reverb | Save Voice (.npz) to create them. Existing files are not overwritten unless overwrite is enabled. The optional migrar_vozes_antigas.bat copies profiles from an older output location without deleting the originals.

Troubleshooting

No module named 'reverb_tts' — install the Python package in the same interpreter used by ComfyUI.

Missing model files — check both the Reverb variant folder and the Higgs codec folder. A .safetensors file by itself is not sufficient.

speed greater than 2 — use the current workflows; older JSONs may have shifted widget values.

No voice in the selector — save a .npz in voices_custom and refresh the frontend.

Dependency conflicts — run python -m pip check using ComfyUI's interpreter. Resolve incompatible dependency requirements carefully; do not indiscriminately upgrade or downgrade packages.

License and responsible use

This repository contains community integration code and example workflows; no model weights are distributed here. The upstream Reverb-TTS project has its own licenses. At the time of writing, the Reverb-TTS model weights are licensed CC BY-NC 4.0 (non-commercial), while upstream code licensing is separate. The present repository declares Apache-2.0 for its own integration code (see LICENSE); the license for model weights remains separate.

Use reference voices only with appropriate consent or authorization, and review the upstream licenses before commercial use.


<a id="português"></a>

🇧🇷 Português

ComfyUI-Reverb-TTS é uma integração experimental do Reverb-TTS ao ComfyUI, com geração de voz local, clonagem por referência, perfis de voz reutilizáveis em .npz e exportação WAV com player integrado.

Recursos principais

  • Carregamento local, sem download automático, das variantes FP32, FP16, INT8 e INT4.
  • Clonagem de voz a partir de áudio e transcrição exata.
  • Vozes oficiais predefinidas armazenadas em JSON.
  • Biblioteca de vozes clonadas .npz com seletor.
  • Geração de fala em português (pt) e outros idiomas.
  • WAV PCM 24-bit, PCM 16-bit ou float, com prévia dentro do node.

Instalação rápida

  1. Clone o repositório dentro de ComfyUI/custom_nodes/.
  2. Instale requirements.txt no Python que executa seu ComfyUI, após conferir conflitos de dependências com --dry-run.
  3. Baixe manualmente os pesos do Reverb, o tokenizer e o codec Higgs para a estrutura explicada acima.
  4. Reinicie o ComfyUI e abra um dos cinco workflows públicos.

Pasta principal configurada nesta versão para Windows:

C:\AI\ComfyUI-Data\models\reverb_tts\

Na v0.3.2, o node Reverb | Load Model (Local) aceita o campo opcional models_directory para informar outra pasta de modelos. Também é possível usar a variável de ambiente REVERB_TTS_MODELS_DIR. Se não houver caminho configurado, o sistema preserva o caminho Windows acima (quando existe) e, caso contrário, procura ComfyUI/models/reverb_tts.

Para que serve cada workflow?

| Nº | Workflow | Resultado | |---|---|---| | 01 | Clonar e gerar | Áudio WAV com voz clonada | | 02 | Voz predefinida | Narração com voz oficial | | 03 | Criar e salvar voz | Perfil de voz .npz | | 04 | Carregar voz e gerar | WAV a partir de perfil salvo | | 05 | Clonar, salvar e gerar | Perfil .npz + WAV |

Biblioteca de vozes: C:\AI\ComfyUI-Data\models\reverb_tts\voices_custom\. Novas vozes podem exigir atualizar a página (F5) para aparecer no seletor.

Atenção: a integração ainda é experimental. Os testes relatados anteriormente foram de estrutura e código; a geração real e a compatibilidade com todos os ambientes precisam ser validadas. Os pesos Reverb-TTS são CC BY-NC 4.0, com restrições a uso comercial. Use vozes de outras pessoas somente com autorização.

v0.3.2 — Custom model path / Caminho personalizado

English: The model loader now supports an optional models_directory input. Leave it blank to use REVERB_TTS_MODELS_DIR if set, otherwise the existing legacy Windows directory (C:\\AI\\ComfyUI-Data\\models\\reverb_tts) when present, otherwise ComfyUI/models/reverb_tts. No models are downloaded automatically. The directory must contain both Reverb-TTS and higgs-audio-v2-tokenizer with their expected contents.

Important: This setting applies to model loading. In the current version, the .npz voice library still uses the default/environment-resolved root, so set REVERB_TTS_MODELS_DIR as well if you move your entire library, or keep voices_custom in its existing location. Existing workflows remain compatible.

Português: O carregador aceita o campo opcional models_directory. Se ficar vazio, usa a variável REVERB_TTS_MODELS_DIR quando definida; senão, mantém a pasta Windows já utilizada (quando existir) e, por último, ComfyUI/models/reverb_tts. Não há download automático. Atenção: a biblioteca de vozes .npz ainda utiliza o diretório padrão/definido pela variável de ambiente; o campo do node não muda a pasta dos perfis.

Dependency compatibility / Compatibilidade de dependências

The package reverb-tts==1.0.0 can introduce conflicting dependency requirements in an existing ComfyUI installation. Run python -m pip install --dry-run -r requirements.txt before installation, and python -m pip check afterwards. Do not assume all custom nodes will remain compatible. No dependency upgrades are required solely to pull this v0.3.2 code change.

O pacote reverb-tts==1.0.0 pode produzir conflitos em instalações com outros custom nodes. Verifique o --dry-run antes de instalar e execute pip check depois. Não é necessário reinstalar dependências apenas para atualizar o código desta versão.

v0.3.3 — Native inline WAV player / Player WAV integrado

Fix: The Reverb | Save WAV node now explicitly declares ComfyUI's AUDIO_UI widget so the frontend creates the native inline audio player. The output still returns ui.audio metadata, retains the saved_path output and saves lossless WAV in the chosen subtype. A frontend widget was missing before, so returning ui.audio alone did not produce a preview.

After updating: completely restart ComfyUI, refresh the browser, remove the old Save WAV node and add a fresh Reverb | Save WAV if the player does not immediately appear, reconnect its audio input, then run the workflow. The player is populated after a new execution; previously generated audio does not automatically reappear. No extra Python dependencies or model downloads are needed. This change is based on ComfyUI's built-in AUDIO_UI widget interface; actual frontend behavior should still be confirmed on the user's installation.

Português: O node Reverb | Save WAV passa a declarar AUDIO_UI, necessário para renderizar o player nativo no frontend. Reinicie o ComfyUI, atualize a página e, se necessário, apague apenas o node antigo de salvamento, adicione o novo e reconecte a entrada audio. Execute novamente para preencher o player. Não é necessário baixar pesos nem reinstalar bibliotecas. A interface precisa ser verificada no ComfyUI real.

v0.3.4 — Installed-model dropdown / Lista de modelos instalados

English: Reverb | Load Model (Local) now includes the optional installed_model dropdown. It detects locally installed Reverb variants (FP32, FP16, INT8, INT4) from the configured environment directory, the legacy Windows folder and the ComfyUI models folder. Only variants with both weights/config and the required codec/tokenizer are shown. No downloads occur. Select Use variant / custom path to retain the previous variant and models_directory behavior. The new field is appended after existing widgets for compatibility. Restart ComfyUI and recreate the loader node if its dropdown does not appear; refresh the page after adding model files.

Português: O node Reverb | Load Model (Local) agora tem o menu installed_model, que mostra as variantes instaladas localmente com os arquivos necessários. Exemplo: Windows / fp32. Se escolher Use variant / custom path, ele continua usando os controles antigos variant e models_directory. Não baixa pesos automaticamente. Reinicie o ComfyUI e recrie apenas o node de carregamento se o menu não aparecer. A interface ainda precisa ser validada em uma instalação real.

v0.3.5 — Simplified Load Model / Carregador simplificado

English: The model loader now uses a single model dropdown listing only complete local model variants found in supported directories (for example Auto / fp32, Windows / int8). The redundant variant field and the per-node models_directory field were removed. device remains available, while force_fp32_compute is an optional advanced checkbox defaulting to false. An alternative models directory can still be configured globally using the REVERB_TTS_MODELS_DIR environment variable before launching ComfyUI. Model weights are never downloaded automatically.

The five public workflow JSON files (01–05) were migrated to Auto / fp32 and will work when an FP32 model is installed in one of the scanned folders. If only another precision is installed, select it from the dropdown. Existing saved workflows using the old loader widget layout should be reloaded using the updated examples or have the loader node replaced. Restart ComfyUI after updating and refresh to rescan model folders.

Português: O node Reverb | Load Model (Local) tem agora um único menu model, com os modelos instalados detectados automaticamente. Os campos variant e models_directory foram retirados da interface. Permanecem device e a opção avançada force_fp32_compute, desativada por padrão. Para usar uma pasta personalizada, defina REVERB_TTS_MODELS_DIR antes de abrir o ComfyUI. Os cinco workflows públicos (01–05) foram adaptados para Auto / fp32. Atualize os workflows e, se necessário, recrie o node carregador antigo. Nenhum modelo é baixado automaticamente.

v0.3.6 — Deduplicated model selector / Seletor sem duplicatas

The model dropdown lists one Auto / <variant> entry for each detected precision. A location-specific entry is displayed only if that variant is also installed in another distinct models directory. The existing Auto / fp32 workflow selections remain compatible. No weights or dependencies change.

O menu de modelos não mostra mais Auto / fp32 e Windows / fp32 quando ambos apontam para os mesmos pesos. Se houver outra instalação realmente distinta da mesma variante, essa segunda pasta poderá ser selecionada separadamente. Reinicie o ComfyUI e atualize o node após fazer o git pull.

v0.3.7 — Fix Save WAV missing connection

English: Fixed the Reverb | Save WAV node incorrectly declaring its audioUI preview widget as a mandatory input. The frontend was asking for a connection to the UI widget and blocking execution. The preview widget is now in the optional inputs section, while the real audio input remains required. WAV saving and ui.audio preview metadata are unchanged. Restart ComfyUI and refresh the workflow; recreate the Save WAV node if it retains the old required slot. Verify the inline player after a fresh execution.

Português: Corrigido o erro Missing connection: Reverb | Save WAV – audioUI. audioUI é um widget de interface, não uma entrada que precisa ser conectada. Agora está declarado como entrada opcional. A entrada audio continua obrigatória. Reinicie o ComfyUI, atualize a página e, se necessário, substitua apenas o node Save WAV por um novo antes de executar novamente. O player deve ser validado na instalação real.

v0.3.8 — Voice dropdowns and speed widget fix

  • Reverb | Preset Voice: preset_name is now a native dropdown populated by locally installed Reverb-TTS/voices/*.json files discovered through model directories. No download, no typed filenames. New JSON voices appear after frontend refresh or restarting ComfyUI.
  • Reverb | Voice Library (NPZ Selector): its existing dropdown lists voices_custom/*.npz profile names; improve tooltip, and refresh ComfyUI after saving new voices. The library displays an empty-library placeholder if no profiles exist. Workflow 04 no longer stores a fixed profile name; the dropdown defaults to the user's available profiles.
  • Reverb | Generate Speech: speed already has backend default 1.0; the shipped workflow JSON files omitted the dynamically inserted control after generate widget after seed, so the speed position was shifted to 8.00. All five workflows that generate speech now explicitly serialize the seed control as fixed and speed as 1.0.
  • Testing: regression tests now validate preset JSON discovery, missing preset behavior, NPZ profile discovery, speed bounds/default, JSON widget positions, WAV saving and loader selection. The regression suite runs in GitHub Actions before release; no Registry publication is triggered automatically.

Português: O seletor Reverb | Preset Voice agora lista os JSONs de vozes oficiais instalados localmente; a biblioteca NPZ também lista as vozes salvas em menu. Nos workflows, speed foi corrigido para 1.00 após o controle do seed. Os modelos continuam locais e não são baixados automaticamente.

v0.3.9 — Removed two redundant nodes

The Reverb | Clear Model Cache output node was removed from the node registry. Existing model loading, speech generation and voice-library functionality remain unchanged. No public example workflow uses the removed node. Note: ComfyUI's own memory-management controls may not clear the extension's Python-level _MODEL_CACHE, which remains an internal optimization; restarting ComfyUI clears that cache if needed.

Português: O node Reverb | Clear Model Cache foi retirado da interface por ser redundante com o gerenciamento de memória do ComfyUI. Nenhum dos cinco workflows públicos utiliza esse node. O cache Python interno continua ativo enquanto o ComfyUI estiver em execução; para liberá-lo completamente, reinicie o ComfyUI.

The Reverb | Load Voice by Name (.npz) node was also removed because Reverb | Voice Library (NPZ Selector) already lets users choose saved profiles from a dropdown. Voice Library now loads the selected NPZ directly, without relying on the removed node. Existing personal workflows using the old Load Voice by Name node must be updated to use Voice Library; the five public workflows remain available.

Português: O node Reverb | Load Voice by Name (.npz) também foi removido. O Reverb | Voice Library (NPZ Selector) já permite escolher a voz numa lista e agora carrega o NPZ diretamente. Workflows particulares que usarem o node antigo devem substituí-lo pelo seletor. Não removemos nenhum perfil de voz salvo.