HojoAI

The AI stack for intelligent hardware.
AgenticOS｜ASR｜TTS｜Memory
Build with Hojo：

[HojoAI @hojoHQ]

Build and ship agentic devices with one AI stack.

Hojo gives edge AI developers everything from Models and Memory to Agent Runtime and Device Access.

Make real devices listen, remember, reason and act.

GitHub: https://t.co/q8H1krjyVr
Hugging Face: https://t.co/g38ngfwz1n

[HojoAI @hojoHQ]

@wngyn9472436216 我们的目标就是在各种复杂环境都能够提供更加鲁棒的识别效果。

[HojoAI @hojoHQ]

Clean audio is easy. Real conversations aren’t.
People switch languages halfway through a sentence. They change their minds, speak over traffic noise, use bad microphones — and still expect the model to keep up.
So we want to test the messy stuff.
Send us a short, non-sensitive audio sample or tell us about a difficult voice scenario. We’ll test it with Hojo-ASR-Multi-V1 and share what works, what doesn’t, and where the model still needs improvement.
What should we test first?
Model: https://t.co/hDRzYy7Uo5

[HojoAI @hojoHQ]

@vivekw_eth 通常模型对于不同的非语音的噪声都有一定的鲁棒性，最难的是语音类的噪声

[HojoAI @hojoHQ]

@froggyoxyz_ 目前对于语音识别来说，干扰人声确实还是比较有挑战的，这种问题可能可以采用针对特定目标人的语音识别模型，也就是让模型只识别你自己的声音，这项工作我们也有涉及，希望很快也可以开放一个这样的模型出来

[HojoAI @hojoHQ]

@li9292 @grok 模型已经在Hugging Face上公开了

[HojoAI @hojoHQ]

@realdoctorwu 针对code-switching和background noise都有优化。对于mid-sentence corrections，目前这个是整句识别模型，所以会综合整句的信息对句中的文字进行判断

[HojoAI @hojoHQ]

@TamikoEconpi 我们的模型原生支持多语种自由说，其本身有能力处理code-switching语音，我们也有针对该问题还有噪声问题进行专门的优化。

[HojoAI @hojoHQ]

@lipa_okx 这个方向可能得先等猫语数据集成熟一下 😄 目前 Hojo-ASR 主要处理人类语音，不过猫叫声和环境声音的识别，确实是另一个很有意思的音频任务。

[HojoAI @hojoHQ]

@ShelbyDeal79623 就是想把模型放到不那么理想的条件里看看。口音、距离、噪声、临时改口都会测，也会保留失败案例。只有知道具体在哪里出错，后面的优化才有意义。

[HojoAI @hojoHQ]

@KtAIFeed It really is. “Noise” is also broader than it sounds—a steady fan, traffic, music, and another person speaking can affect ASR very differently. We’re thinking of testing the same sentence against several noise types so the results are easier to compare.

[HojoAI @hojoHQ]

@ponyodong 我们也很想把这个场景测得更细一点：英文专有名词夹在中文里，和说到一半直接切换成英文，其实是两种不同难度。再叠加噪声后，应该会暴露出一些很具体的问题。

[HojoAI @hojoHQ]

@wangyan82A7 这个组合确实够难，而且我们准备把两个变量拆开测：先看中英混说本身，再加入交通声、人声等不同噪声。这样才能看清楚问题究竟出在语言切换，还是声学环境上。后续会把原始音频和识别结果一起发出来。

[HojoAI @hojoHQ]

Exactly. And we don’t want to collapse all of those into one generic “real-world test.” Code-switching, background noise, and poor microphones can cause very different kinds of errors. We’re planning to test them separately and share the raw outputs, including the cases that don’t work well.

[HojoAI @hojoHQ]

Small voice models get interesting when they leave the demo.
Hojo-TTS-Light-40M runs on CPU and supports both Chinese and English — making local voice practical for the products people actually use.
An e-reader?
A game character?
A local assistant?
A creator tool?
Where would you put a voice?
https://t.co/S81hk1VkXI

[HojoAI @hojoHQ]

我们正在测试我们的全双工模型，Hojo-Chat-Duplex

[Google DeepMind @GoogleDeepMind]

We’re introducing Gemini 3.8 Live and 3.8 Live Extended Thinking – our best conversational AI.

The models talk, think, and handle tasks in the background without breaking your flow. 🧵

[HojoAI @hojoHQ]

Voice is becoming the front door to serious AI work. The better the system listens, the less the interface gets in the way.

[Google @Google]

Deep Research 🤝 Gemini Live.

Use your voice to explore a topic — in depth — with Gemini Live's Deep Research integration.

1. Just ask the @GeminiApp to run a Deep Research report on a topic, then feel free to close the chat, lock your screen, or keep chatting about other things.

2. Gemini will work asynchronously in the background and send you a notification when your full research report is ready.

3. Once it is, you can chat through the results, ask follow-up questions, or refine the details.

[HojoAI @hojoHQ]

Can your ASR keep up with a real conversation?

Real conversations don’t stay in one language — or one version of the sentence.
“Let’s meet at 七点半… actually, make it eight.”
Can an ASR model keep the language switch, the time, and the correction — without cleaning up what you actually said?
We’re putting Hojo-ASR through the kind of code-switching people use every day. Drop us a sentence you want tested.

[HojoAI @hojoHQ]

AI hardware needs better ears.

AI is moving closer to the body — into earbuds, glasses, and other always-on devices.
But ambient AI only works when it can hear real life: traffic, accents, interruptions, and mid-sentence corrections.
AI hardware needs better ears.
That’s where Hojo-ASR comes in — helping intelligent hardware turn speech into reliable input in the situations people actually use it.
What voice scenario should we test next?
📷

[HojoAI @hojoHQ]

Hearing clearly is only the start.

You can hear someone perfectly clearly and still get stuck on a local expression. Turning up the volume won’t help.
The Sichuan dialect test shared by @MinLiBuilds brings up a similar question for ASR: how much can you work out from the sound alone?
Regional speech comes with familiar shortcuts, local vocabulary, and pronunciations that don’t neatly match standard Mandarin. People who speak it every day barely notice. A speech recognition system has to work through all of it.
Sometimes, the rest of the sentence is what helps you work out an unclear word. Knowing how people actually use the language matters.
Hojo-ASR-V1 combines acoustic features with a Qwen3 LLM decoder, bringing language knowledge into the recognition process itself. The audio provides evidence of what was said; linguistic context helps the model make sense of that evidence when the sounds are ambiguous.
The transcript still needs to stay faithful to the recording, including the speaker’s dialect expressions and corrections.
That’s why Sichuan dialect is an interesting test. It asks how well a system handles the way people speak at home, with friends, or halfway through a thought.
For someone building a voice-enabled device, that’s a practical concern. Users shouldn’t need to rehearse a sentence in standard Mandarin just to get it transcribed.
Try Hojo-ASR-V1 with the way people actually speak around you.
https://t.co/uAJeKbsCRJ

[HojoAI @hojoHQ]

A smart device should be easy to talk to.
Your hands are full of groceries. You tell your earbuds, “Remind me to leave at seven thirty tomorrow. Actually, make that eight.”
You keep looking for your keys. You expect the reminder to be set for eight.
That’s what people want from AI hardware: to say something naturally and have the device keep up.
With glasses, it might be asking for directions while walking. With earbuds, setting a timer over the kitchen fan. With a robot, asking it to pick something up, then adding, “Wait, leave it there.”
All of these interactions need a reliable way to get spoken words into the system. ASR provides that input. If it misses a correction or a device name, the model handling the request has less to work with.
We see Hojo-ASR as the ears of AI hardware: a voice input layer that developers can build into glasses, earbuds, and robots. Hojo-ASR-Multi-V1 gives teams a starting point for testing those interactions across languages.
Making the whole experience work also takes good microphones, audio processing, and timely responses. Speech recognition belongs in that conversation from the start of product design.
Because once you’ve got both hands full, “Please repeat that” gets old pretty quickly.
https://t.co/hDRzYy7Uo5