Ultravox is a multimodal Speech LLM built around a pretrained LLM (Llama, Gemma, Qwen, etc) and a speech encoder (whisper-large-v3-turbo) 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 v0.6 series, ultravox models are trained on expanded Hindi speech data, resulting in significantly improved speech understanding performance on Hindi and modest degradation on other languages. Additionally, the v0.6 models are also trained on noise datasets for improved noise robustness and the ability to output a special string ((noise)) if the input audio is too noisy or doesn't contain clear speech.

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 LLM (Llama, Gemma, Qwen, etc) backbone as well as the encoder part of whisper-large-v3-turbo.

The multi-modal adapter is trained, the Whisper encoder is fine-tuned, and the LLM is kept frozen.

We use a knowledge-distillation loss where Ultravox is trying to match the logits of the text-based LLM backbone.

The training dataset is a mix of ASR datasets, extended with continuations generated by Llama 3.1 8B, speech translation datasets, and noise datasets.

Supervised speech instruction finetuning via knowledge-distillation. 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 150ms, and a tokens-per-second rate of ~50-100 when using an A100-40GB GPU, all using a text-based LLM (Llama, Gemma, or Qwen) backbone.

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

Evaluations are conducted on covost2 (speech translation measured in BLEU), fleurs and ultravox_calls (speech recognition measured in WER), big bench audio (audio reasoning measured in accuracy), as well as musan and ultravox_unintelligible (noise/unintelligible speech detection measured in recall).