# Nous Research adds streaming speech to speed Hermes Agent voice chats > Source: > Published: 2026-07-29 05:46:11+00:00 [Ryan "Teknium" Teknium (@Teknium)](https://x.com/Teknium), a Nous Research co-founder and its head of post-training, [said on X](https://x.com/teknium/status/2082339029375426914?s=46) Wednesday that Hermes Agent voice chats now stream speech from most API-based text-to-speech backends, bringing the open-source assistant closer to a live conversation. The implementation does not make the underlying language model generate answers faster. It reduces the delay users hear by starting audio playback before Hermes has finished writing its response. That distinction matters in voice products, where silence between a question and the first spoken word can make a capable model feel unresponsive. Teknium helped form [Nous Research](https://nousresearch.com/) as an AI research community on Discord in 2022 before it became a company in 2023. According to [his personal site](https://teknium.io/), he previously worked as an LLM data engineer at Stability AI and on character-focused language models at Digi AI. At Nous, he has led post-training, evaluations and deployment while helping build the Hermes model family. ### How Hermes streams an answer The [Hermes Agent voice documentation](https://hermes-agent.nousresearch.com/docs/user-guide/features/voice-mode/) shows that the system collects incoming text from the model until it has a complete sentence of at least 20 characters. It removes markdown, emoji and hidden reasoning tags, then sends that sentence into the configured text-to-speech provider while the model continues generating the rest of its answer. There are two levels of streaming inside that pipeline. ElevenLabs and OpenAI support chunked PCM audio, allowing Hermes to play raw audio as the provider returns it. Other providers, including the default Edge TTS option, synthesize and play each sentence after it is completed. The first approach offers the shortest path to the first audible word; the second still avoids waiting for the model's entire response. Nous says the same speech pipeline operates in the command-line interface, terminal UI and desktop app. In the desktop implementation, generated text moves through a WebSocket tied to a single reply, keeping model generation and audio playback running at the same time without opening a new connection for every sentence. Hermes also supports barge-in controls. A user can begin talking over a reply, send another message or press the recording key to stop playback. The next prompt tells the model that its prior response was interrupted, giving it context to continue or change direction. The voice system sits inside the broader [open-source Hermes Agent project](https://github.com/NousResearch/hermes-agent), which combines persistent memory, learned skills, browser automation, coding tools and messaging integrations. Users can choose their own model provider rather than being limited to a Nous model, while [Nous Portal](https://portal.nousresearch.com/) sells access to hosted models, tools and cloud deployments through one account. ### Nous is turning Hermes into a business The latency work lands as Nous builds a commercial layer around its open-source agent. [TechCrunch reported on July 13th](https://techcrunch.com/2026/07/13/hermes-agent-maker-nous-research-in-talks-for-new-funding-at-1-5b-valuation/) that Nous was finalizing at least $75 million in financing led by Robot Ventures, with significant participation from Union Square Ventures, at a reported $1.5 billion valuation. The report said Nous had previously raised $70 million from investors including Paradigm, Robot Ventures, North Island Ventures, Delphi Ventures, OSS Capital and Balaji Srinivasan. That reported financing puts pressure on Nous to turn Hermes adoption into a durable hosted product. Voice is useful for that push because it expands the agent beyond typed terminal sessions and messaging bots into hands-free work, desktop assistance and live Discord conversations. The commercial opening is straightforward: keep the agent and provider choices open, then charge users who want Nous to supply the models, speech services, tools and always-on infrastructure. Streaming speech addresses one narrow part of that plan, but it is a consequential one. Users judge a voice agent during the pause after they stop talking. Hermes can now fill that pause sooner, without requiring Nous to control the model generating the answer.