Mosi

Fish Audio is a Palo Alto-based voice AI company building text-to-speech and speech-to-text models with expressive, emotion-aware synthesis across 83+ languages. Started from an open-sourced model trained on a single GPU, it reached ~$21M ARR and 8M+ users within its first year.

Seed · $52M · 2026-07

Deepgram, founded in 2015 in the USA, builds voice AI models and APIs including speech-to-text (STT) and text-to-speech (TTS).

Series C · $130M · 2026-01 · Valuation $1.3B

Nuance Labs, founded in 2025 in Seattle by three former Apple researchers, is building a single full-duplex audiovisual foundation model for face-to-face AI conversation. Instead of chaining separate transcription, language-model, voice-synthesis and animation components, the model ingests live audiovisual signals (tone, gaze, gesture, timing) and streams real-time facial and vocal responses from one model. Target use cases include sales and customer service, coaching, professional training and education (language learning, interview practice). A public research preview is planned for later in 2026 and no commercial product has shipped yet. Note: it is unrelated to Nuance Communications, the speech-recognition company acquired by Microsoft in 2021.

Series A · $60M · 2026-09

Shenzhu Jizhi is a Beijing-based AI company founded in 2025, co-incubated by Zhongguancun Academy and Zhongguancun AI Research Institute, developing PhysBrain, an embodied general intelligence foundation model trained on first-person multimodal data.

$14M · 2026-06

Tripo AI (VAST), founded in 2023 in Beijing, China and San Francisco, USA, develops models and tools for generating, editing, and simulating 3D assets and persistent interactive environments, serving gaming, film, and industrial applications.

Series B · $694M · 2026-09 · Valuation $1.0B

HiDream.ai, founded in 2023 in Beijing, China, is a generative AI developer.

Series C · $208M · 2026-07

Deepslate is a Berlin-based voice AI lab founded in 2024 that builds native speech-to-speech models processing spoken language directly without text conversion, with ~440ms first-audio latency across 27 languages, hosted entirely on EU infrastructure for data-sovereignty-sensitive enterprises such as insurers, energy providers and contact centers.

Seed · 2026-10

Poolside, founded in 2023 in Paris, France, builds a platform that enables developers to write code using AI. It has risen to #2 on the AI coding leaderboard.

Undisclosed · $1.1B · 2024-10

Gradium, founded in 2025 in Paris, France, develops ultra-low-latency audio language AI models and provides real-time voice interaction capabilities.

Seed · $70M · 2025-12

星源智 (Xingyuanzhi), founded around August 2025 in Beijing, develops on-device embodied AI systems including multimodal foundation models, spatial intelligence, and world models for real-world physical interaction.

$139M · 2026-06

Veeda AI builds multimodal foundation world models that simulate physical reality, creating infinitely scalable environments where embodied agents can learn through interaction. Its stated mission is to unlock the scaling moment for Physical AI. The company argues embodied intelligence will not emerge from imitation learning alone, and that the real world is unsafe, expensive and slow as a training ground, so robots must learn in simulated reality. It maintains offices in Toronto, Zurich, the Bay Area and Singapore.

Seed · 2026-08

Smallest.ai, founded in 2024 in Bengaluru, offers an ultra-low-latency voice generation and voice AI model platform targeting enterprise customers.

Series A · $21M · 2026-07