VibeVoice 1.5B

VibeVoice-1.5B is a text-to-speech (TTS) model that generates expressive, long-form, multi-speaker conversational audio from text, such as podcasts or dialogue. It is hosted on gigarouter as a managed, OpenAI-compatible API. The model uses a next-token diffusion framework built on a Qwen2.5-1.5B large language model (LLM), combined with continuous speech tokenizers (acoustic and semantic) operating at an ultra-low frame rate of 7.5 Hz. This architecture enables synthesis of speech up to 90 minutes long with up to 4 distinct speakers, capturing natural turn-taking and conversational flow. The tokenizer achieves an 80× improvement in data compression over Encodec while maintaining comparable audio fidelity.

Key Capabilities

• Generates long-form, multi-speaker conversational audio (e.g., podcasts) from text input.
• Supports up to 4 distinct speakers in a single generation.
• Context length of 64,000 tokens, enabling synthesis of up to 90 minutes of continuous speech.
• Built on a Qwen2.5-1.5B LLM with a diffusion-based decoding head (4 layers, ~123M parameters).
• Continuous speech tokenizers (acoustic and semantic) operate at 7.5 Hz, achieving 80× better data compression than Encodec while maintaining comparable audio fidelity.

Model Variants

Model	Context Length	Generation Length
VibeVoice-0.5B-Streaming	—	—
VibeVoice-1.5B	64K tokens	~90 minutes
VibeVoice-Large	32K tokens	~45 minutes

Architecture

The model uses a transformer-based LLM (Qwen2.5-1.5B) integrated with acoustic and semantic tokenizers and a diffusion-based decoding head. The acoustic tokenizer, based on a σ-VAE variant, achieves 3200× downsampling from 24 kHz input. The diffusion head (4 layers, ~123M parameters) predicts acoustic VAE features using a Denoising Diffusion Probabilistic Models (DDPM) process with Classifier-Free Guidance (CFG) and DPM-Solver during inference.

Benchmarks and Research Context

VibeVoice was accepted as an Oral at ICLR 2026. The underlying LatentLM framework, which VibeVoice builds upon, outperforms the state-of-the-art VALL-E 2 model in speaker similarity and robustness while requiring 10× fewer decoding steps. The TTS code repository was disabled on 2025-09-05 due to observed out-of-scope use; the model is intended for research purposes only.

Additional Information

The model is trained on English and Chinese data only. It does not handle overlapping speech, background noise, music, or other non-speech audio. Generated audio includes an audible AI disclaimer and an imperceptible watermark for provenance verification. For further details, refer to the technical report and project page.