O Que É AUDIO

July 29 – According to official sources, AI voice company Fish Audio announced the completion of a $52 million seed funding round, with specific investors undisclosed. The company also officially launched its next-generation voice generation model, S2.1 Pro. Fish Audio says S2.1 Pro can clone voices using just 5 seconds of audio samples, generating speech approximately twice as fast as Cartesia at about one-sixth the cost of ElevenLabs. Additionally, the model supports fine-grained control over word-level emotion, intonation, and speaking speed, making it one of the most expressive voice models currently available. Several AI companies, including HeyGen, LiveKit, Retell, Sanas, and OpenArt, are already using Fish Audio models in production. The company stated that if an enterprise adopts its voice AI service and fails to reduce costs by at least 50%, Fish Audio will provide one year of free service. To celebrate its first anniversary, Fish Audio is also offering users a free one-month trial of S2.1 Pro.

July 28, according to TechCrunch, Fish Audio announced on Tuesday that it has raised $50 million in seed funding. The round was led by Coreline Ventures and Capital Today, with participation from 359 Capital, Parable, Play Time, Alphalist Partners, Bayhouse Ventures, Carya Venture Partners, and HF0. Fish Audio initially started as a small project by Shijia Liao, a former researcher at Nvidia. Dissatisfied with the lack of expressive synthetic voice technology on the market, he trained a voice generation model using a single GPU and open-sourced it. Today, the Fish Speech project repository has garnered over 31,000 stars on GitHub and is widely used by independent developers, game designers, and content creators. Over the past year, Fish Audio has released five models, including four voice generation models and one speech-to-text model. The company has open-sourced three of its voice generation models, but its latest S2.1 Pro model is only available via a paid API. Since its launch last year, the startup's open-source and hosted versions have attracted over 8 million users, and its current annual recurring revenue (ARR) has reached $21 million.

On April 29, NVIDIA officially launched the Nemotron 3 Nano Omni, a new member of the Nemotron 3 series, which integrates unified multimodal inference into a single efficient open-source model. NVIDIA stated that agentic systems typically require reasoning in a single perception-to-action loop across screens, documents, audio, video, and text, but still rely on fragmented model chains—separate technology stacks for vision, audio, and text. This increases the number of inference hops and orchestration complexity, raising inference costs while weakening the consistency of cross-modal context. The Nemotron 3 Nano Omni aims to replace this fragmented visual-language-audio technology stack, serving as a multimodal perception and context sub-agent within agentic systems. In terms of accuracy, the Nemotron 3 Nano Omni has achieved leading results on document intelligence leaderboards, while also leading in video and audio understanding benchmarks. On the open industry benchmark MediaPerf for evaluating video understanding models, the Nemotron 3 Nano Omni achieved the highest throughput in every task and recorded the lowest inference cost in video-level annotation tasks. In terms of performance, under a fixed per-user interaction threshold, for video inference, the Nemotron 3 Nano Omni maintains a higher total system throughput, achieving up to approximately 9.2 times the effective system capacity compared to other open-source omni models; for multi-document inference, it can achieve up to approximately 7.4 times the effective system capacity. NVIDIA stated that this model is designed to replace traditional multi-model stitching architectures, reducing inference complexity and costs, and promoting the application of multimodal AI in scenarios such as finance, healthcare, scientific research, and media.