Ultravox is a multimodal Speech LLM built around a pretrained Llama3.1-8B-Instruct and Whisper-small backbone.
See https://ultravox.ai for the GitHub repo and more information.

Ultravox is a multimodal model that can consume both speech and text as input (e.g., a text system prompt and voice user message). The input to the model is given as a text prompt with a special <|audio|> pseudo-token, and the model processor will replace this magic token with embeddings derived from the input audio. Using the merged embeddings as input, the model will then generate output text as usual.

In a future revision of Ultravox, we plan to expand the token vocabulary to support generation of semantic and acoustic audio tokens, which can then be fed to a vocoder to produce voice output. No preference tuning has been applied to this revision of the model.

Think of the model as an LLM that can also hear and understand speech. As such, it can be used as a voice agent, and also to do speech-to-speech translation, analysis of spoken audio, etc.

To use the model, try the following:

The model uses a pre-trained Llama3.1-8B-Instruct backbone as well as the encoder part of Whisper-small.

Only the multi-modal adapter is trained, while Whisper encoder and Llama are kept frozen. We use a knowledge-distillation loss where Ultravox is trying to match the logits of the text-based Llama backbone.

Training dataset is a mix of ASR datasets, extended by adding a "continuation" generated by Llama 3.1 8B.

Supervised speech to audio finetuning. For more info, see training code in Ultravox repo.

The current version of Ultravox, when invoked with audio content, has a time-to-first-token (TTFT) of approximately 200ms, and a tokens-per-second rate of ~50-100 when using an A100-40GB GPU, all using a Llama 3.1 8B backbone.

Check out the audio tab on TheFastest.ai for daily benchmarks and a comparison with other existing models.