# VibeVoice Community Fork for Long-Form Conversational Speech

> VibeVoice is a Python speech-generation framework for expressive, multi-speaker audio, including long podcast-style conversations. This community fork preserves the codebase after the official repository was removed and adds unofficial training and fine-tuning support, with model choice determining the trade-off between duration, speaker count, and real-time use.

**vibevoice-community/VibeVoice** — VibeVoice: Expressive, longform conversational speech synthesis. (Community fork)

- Repository: https://github.com/vibevoice-community/VibeVoice
- Website: https://discord.gg/ZDEYTTRxWG
- Stars: 1,635 · Forks: 780
- Language: Python
- License: MIT
- Published: 2026-10-16 · Updated: 2026-10-16 · Language: en
- Canonical page: https://hysenlabs.com/projects/vibevoice-community-vibevoice

## What VibeVoice is designed to generate

VibeVoice targets a specific gap in conventional text-to-speech: generating expressive, long-form conversations rather than isolated utterances with one stable narrator. The documented use case is podcast-style audio, with the model able to synthesize up to 90 minutes and support up to four distinct speakers. That makes it relevant to teams building narrated dialogue, conversational demonstrations, fictional interviews, and multi-voice document or topic productions.

The project is not limited to the standard 1.5B model. The model table includes VibeVoice-Streaming-0.5B for real-time, single-speaker synthesis, VibeVoice-1.5B for approximately 90-minute generation with up to four speakers, and VibeVoice-Large, or 7B, for approximately 45-minute generation with up to four speakers. The fork also lists Hugging Face Transformers variants, including VibeVoice-1.5B-hf and VibeVoice-7B-hf.

The scope is audio generation, not a complete podcast publishing system. The roadmap still leaves VibePod, described as an end-to-end solution for creating podcasts from documents, webpages, or a topic, unchecked. Users who need script preparation, asset management, editing, and export as one workflow should not treat the current repository as that product.

## The token and diffusion architecture

The README identifies two continuous speech tokenizers, acoustic and semantic, operating at a 7.5 Hz frame rate. The stated purpose is to preserve audio fidelity while reducing the computational cost of processing long sequences. That design is directly connected to the project's long-duration target. A lower-frame-rate representation can make a long dialogue more practical than a system that predicts fine-grained speech units at a much higher rate.

VibeVoice uses a next-token diffusion framework. According to the repository description, an LLM handles textual context and dialogue flow, while a diffusion head generates acoustic detail. The separation gives the model two related jobs: interpreting who should speak, what should be said, and how the conversation progresses, then producing the acoustic representation of the speech.

The model table also exposes an important resource trade-off. VibeVoice-1.5B has a 64K context length and an estimated generation length of about 90 minutes. The 7B model has a 32K context length and an estimated generation length of about 45 minutes. Choosing the larger model does not simply mean choosing a better default. It changes the context budget and may impose greater hardware demands. The README notes that only VibeVoice-1.5B is supported in its Colab notebook because of GPU memory limitations.

## Model selection for duration, speakers, and latency

The model comparison in the README gives a practical selection rule. Use VibeVoice-Streaming-0.5B when real-time, low-latency speech is the priority and one speaker is sufficient. It uses pre-computed voice embeddings stored as `.pt` files for fast inference. This is a different operating mode from the long-form 1.5B and 7B models, which support voice cloning from audio samples and multi-speaker content.

Use VibeVoice-1.5B for the longest documented generation window. Its approximately 90-minute output and 64K context length make it the natural candidate for a complete conversation or podcast-style script. The 7B variant supports up to four speakers as well, but its listed generation length is approximately 45 minutes. The larger parameter count should therefore be evaluated against the need for model capacity rather than assumed to be the better choice for every script.

For direct file inference, the repository supplies scripts under `demo/`, including `demo/inference_from_file.py` and `demo/streaming_inference_from_file.py`. The README also documents a Gradio interface. These entry points make local experimentation possible without requiring users to assemble a generation pipeline themselves, although the repository does not document a single universal command for every supported model.

## Installing the Python package and running the Gradio demo

The README gives the basic installation sequence. Clone the community repository, enter its directory, and install the package in editable mode with `uv pip install -e .`. The project requires Python 3.9 or later. Its `pyproject.toml` pins `accelerate==1.6.0`, `transformers==4.51.3`, `datasets==3.5.0`, and `gradio==5.50.0`, so dependency compatibility is part of the setup rather than an incidental detail.

The first documented interactive use is the Gradio demo for the 1.5B model:

```bash
git clone https://github.com/vibevoice-community/VibeVoice.git
cd VibeVoice/
uv pip install -e .
```

After installation, launch the interface with the model path used by the README:

```bash
python demo/gradio_demo.py --model_path vibevoice/VibeVoice-1.5B --share
```

The command starts the demo and passes `--share` to expose it through a shareable interface, subject to the environment and Gradio behavior. The README also lists a 7B alternative, a `--checkpoint_path` option for a fine-tuned adapter, and an in-app `Disable voice cloning` setting under Advanced Settings. Those options should be selected deliberately, because voice conditioning changes the inputs required for a synthesis session.

## File-based inference and fine-tuned adapters

For a local first synthesis, the README documents `demo/inference_from_file.py`. A single-speaker example uses the 7B model and the `Alice` speaker name. A two-speaker example uses `Alice` and `Frank` with a different text example. The model path, text path, and speaker names are supplied as command-line arguments, so the script is the point where text, speaker assignment, and model selection meet.

```bash
python demo/inference_from_file.py --model_path vibevoice/VibeVoice-7B --txt_path demo/text_examples/1p_abs.txt --speaker_names Alice
```

The README also documents loading a fine-tuned LoRA checkpoint:

```bash

```

The fork's stated additional functionality includes unofficial training and fine-tuning implementations. `FINETUNING.md` is present as a top-level repository entry, while `peft` is included as a dependency. The documentation indicates adaptation to a new language or a new voice, but the excerpt does not provide a complete training command or hardware matrix. Fine-tuning should therefore be treated as a documented extension with its own separate instructions, not as an automatic part of installation.

## Known synthesis constraints and maintenance boundaries

The README explicitly warns about occasional instability when synthesizing Chinese speech. Its suggested mitigations are English punctuation for Chinese text, preferably commas and periods, use of the Large model variant, and chunking text into multiple speaker turns when the generated voice speaks too quickly. These are workarounds documented by the project, not guarantees that every Chinese script will behave consistently.

The project also has dependency constraints. The `pyproject.toml` says it develops against `transformers==4.51.3` and warns that later versions may not be compatible. The README's update entry says VibeVoice is part of Transformers and points to 1.5B and 7B model cards for examples including batched inference and `torch.compile`. However, the roadmap leaves Transformers integration for the real-time or streaming model unchecked. A streaming deployment should not assume that the same integration status applies to every model variant.

This repository is not archived, and the last push was on 2026-08-29. There are no GitHub releases. The README also calls this a community-maintained fork created after the official repository was removed and restored without code. That history matters when evaluating upgrade expectations: model cards, dependency pins, and community integrations may change independently of a tagged release.

The MIT license permits use and modification under its terms, subject to the license file's conditions. The repository is not a hosted speech service, and the documentation does not promise availability, moderation, or support for generated audio. Users must separately consider consent and rights for voice samples and the intended distribution of generated speech.

## VibeVoice compared with ordinary single-voice TTS pipelines

A conventional TTS pipeline usually maps text to one speaker's waveform and then relies on application-level concatenation for longer material. VibeVoice instead presents a model designed around dialogue context, speaker turns, long-form generation, and acoustic detail produced through a diffusion head. The practical difference is not merely a larger waveform. Speaker assignment is represented in the generation task, and the listed long-form models are intended to preserve conversational flow over substantially longer output than a typical short utterance workflow.

The real-time 0.5B model narrows that distinction. It is designed for one speaker, real-time generation, and pre-computed voice embeddings, so it resembles a latency-oriented TTS component more than a podcast engine. VibeVoice becomes the more relevant option when the core requirement is a multi-person conversation, a long script, or generation from speaker-conditioned audio.

The trade-off is that model selection is not independent from workload. The 1.5B model provides the longest listed generation window, the 7B model offers a larger model but a shorter listed generation window, and the streaming model changes the speaker and conditioning model. A team should compare context length, generation length, speaker count, and inference mode before choosing an alternative.

## Conclusion

Adopt VibeVoice when the requirement is long-form conversational speech with multiple speakers and the available model fits the required duration and latency. Do not choose it for unattended short-form TTS without checking Chinese-text stability, model memory needs, and the differences between the 1.5B, 7B, and streaming variants. First verify the selected model card, GPU requirements, voice-conditioning inputs, and whether the required Transformers integration is available.

## FAQ

### How do I install VibeVoice?

Clone `vibevoice-community/VibeVoice`, enter the repository directory, and run `uv pip install -e .`. The project metadata requires Python 3.9 or later and pins several dependency versions.

### Which VibeVoice model should I use for multi-speaker podcasts?

The model table lists VibeVoice-1.5B and VibeVoice-7B as supporting up to four speakers. VibeVoice-1.5B is associated with approximately 90 minutes of generation, while VibeVoice-7B is associated with approximately 45 minutes.

### Does VibeVoice support real-time speech and voice cloning?

VibeVoice-Streaming-0.5B is documented for real-time, low-latency TTS with one speaker and pre-computed `.pt` voice embeddings. The long-form 1.5B and 7B models are documented as supporting voice cloning from audio samples, while the roadmap says Transformers integration for the streaming model is not yet complete.

### What should I check if Chinese synthesis sounds unstable?

The README recommends using English punctuation, preferably commas and periods, trying the Large model variant, and chunking text with multiple speaker turns if the voice speaks too quickly. These are the project's documented mitigations for occasional Chinese speech instability.

## Sources

- [Issues](https://github.com/vibevoice-community/VibeVoice/issues)
- [License: MIT](https://github.com/vibevoice-community/VibeVoice/blob/main/LICENSE)
- [Project website](https://discord.gg/ZDEYTTRxWG)
- [README](https://github.com/vibevoice-community/VibeVoice/blob/main/README.md)
- [vibevoice-community/VibeVoice on GitHub](https://github.com/vibevoice-community/VibeVoice)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/vibevoice-community-vibevoice