Boson AI

Boson AI

Boson AI is an audio foundation model company behind the Higgs Audio family of models for text-to-speech, speech-to-text and audio understanding. Boson AI was founded in 2023 by Dr. Alex Smola and Dr. Mu Li, both former AWS AI leaders and co-authors of the widely used textbook Dive into Deep Learning. The flagship Higgs Audio v2 was pretrained on more than 10 million hours of audio, generating expressive speech with control over emotion and multiple speakers. Newer releases add Higgs Audio 2.5 for production voice and Higgs Audio 3.0, a speech-to-text model spanning 94 languages. Boson AI ships open weights on Hugging Face and offers aligned models for enterprise use.

Production credibility: Boson AI is based in Santa Clara, California and was founded in 2023 by Dr. Alex Smola and Dr. Mu Li, both former AWS AI executives and co-authors of the open-source deep learning textbook Dive into Deep Learning. Its Higgs Audio v2 model was pretrained on over 10 million hours of audio plus text data and released with open weights (a 3B-parameter generation model on Hugging Face). On the EmergentTTS-Eval benchmark, Higgs Audio v2 reported win rates of 75.7% and 55.7% over gpt-4o-mini-tts on the Emotions and Questions categories, with strong results on Seed-TTS Eval and the Emotional Speech Dataset. Higgs Audio 3.0, a speech-to-text foundation model covering 94 languages, reported lower word error rate than Whisper-large-v3 on tested sets (for example 1.55 vs 2.10 WER on LibriSpeech clean). Boson AI also published the EmergentTTS-Eval benchmark (accepted at NeurIPS 2025) and partnered with Eigen AI to ship a smaller 1B-parameter build of Higgs Audio 2.5. Funding details have not been publicly disclosed.

Key Features

• Higgs Audio v2: expressive text-to-speech foundation model pretrained on 10M+ hours of audio
• Higgs Audio 3.0: speech-to-text (ASR) model supporting 94 languages with language detection
• Higgs Audio 2.5: production-focused voice generation with reduced latency
• Audio understanding with real-time reasoning over sentiment and semantics
• Open weights published on Hugging Face and code on GitHub
• Multi-speaker and emotion-controllable speech synthesis
• Published the EmergentTTS-Eval benchmark (NeurIPS 2025) for expressive TTS
• Aligned foundation models and custom solutions for enterprise via direct contact

Ideal Use Case

Developers and enterprises use Boson AI's Higgs Audio models to add expressive text-to-speech, multilingual transcription and real-time audio understanding to voice agents, assistants and media tools, either via open weights or aligned enterprise builds.

How Boson AI differentiates

Against ElevenLabs and Cartesia, Boson AI's distinction is that it ships open-weight audio foundation models rather than a closed voice API, so teams can self-host Higgs Audio and inspect the model. Unlike ElevenLabs, which is a polished hosted product with a large voice library, Boson AI sits closer to the research and model layer, founded by Dive into Deep Learning authors Alex Smola and Mu Li. Compared with Cartesia's latency-tuned streaming voices, Boson covers a wider stack, from expressive TTS to 94-language speech-to-text and audio understanding in one family. The trade-off is that open models require more engineering to deploy, and Boson's hosted product and pricing are less turnkey than ElevenLabs for non-technical buyers.

FAQ

Q: Who founded Boson AI? A: Boson AI was founded in 2023 by Dr. Alex Smola and Dr. Mu Li, both former AWS AI leaders and co-authors of the open-source textbook Dive into Deep Learning. The company is based in Santa Clara, California.

Q: How much funding has Boson AI raised? A: Boson AI has not publicly disclosed its funding amount or investors. It operates as a privately held audio foundation model company founded by Alex Smola and Mu Li, releasing the Higgs Audio models with open weights.

Q: What is Higgs Audio? A: Higgs Audio is Boson AI's family of audio foundation models. Higgs Audio v2 and 2.5 handle expressive text-to-speech, and Higgs Audio 3.0 is a speech-to-text model covering 94 languages. Several versions are released with open weights on Hugging Face.

Q: Boson AI vs ElevenLabs: what's the difference? A: ElevenLabs is a closed, hosted voice API with a large voice library and turnkey product. Boson AI publishes open-weight Higgs Audio foundation models you can self-host and fine-tune, covering TTS, 94-language speech-to-text and audio understanding, but with more engineering required to deploy.

Q: Is Higgs Audio open source? A: Boson AI releases open weights for several Higgs Audio models on Hugging Face, including a 3B-parameter generation model, along with code on GitHub. It also offers aligned foundation models and custom solutions for enterprises on request.

tl;dr

Boson AI builds the Higgs Audio foundation models for text-to-speech, 94-language speech-to-text and audio understanding, with open weights on Hugging Face. Founded in 2023 by Dive into Deep Learning authors Alex Smola and Mu Li, it reports benchmark wins over gpt-4o-mini-tts and Whisper-large-v3.

Related

Looking for more options? Browse the AI/ML Models directory or read our best AI models listicle. Boson AI is also tracked on Crunchbase.