Hojo-ASR-V1
- Task
- Speech recognition
- Architecture
- LLM decoder
- Languages
- Chinese and English, including Cantonese and Sichuanese
- License
- Apache-2.0
Between hardware and the real world: Hojo’s acoustic models.
Hojo’s in-house speech models are acoustic models, the layer through which hardware talks with the real world. Hojo-ASR turns what it hears into text; Hojo-TTS-Light speaks the reply. All four are open source under Apache-2.0.
Speech recognition: best result, second worldwide.
Hojo ASR uses a large language model as its decoder and comes in a Chinese–English edition and a multilingual one. Its best result on the Open ASR Leaderboard was second worldwide, in June 2026.
Open ASR Leaderboard · June 2026
Hojo ASR’s best result: second worldwide ↗Hojo-ASR-Multi-V1
- Task
- Speech recognition
- Architecture
- LLM decoder
- Languages
- German, French, Italian, Portuguese, Spanish
- License
- Apache-2.0
Meetings longer than five hours, transcribed live.
Hojo’s in-house ASR can transcribe meetings that run past five hours, in real time.
Hojo ASR
Real-time transcription for meetings over five hours.
Lightweight speech synthesis that runs on a CPU.
Hojo-TTS-Light is a family of lightweight speech generation models for Chinese and English, in two sizes: 80 million and 40 million parameters.
Hojo-TTS-Light-40M
40 million parameters, with an ONNX version that runs on an ordinary CPU.
Hojo-TTS-Light
- Task
- Speech synthesis
- Parameters
- 80M
- Languages
- Chinese, English
- Highlights
- Voice cloning
- License
- Apache-2.0
Hojo-TTS-Light-40M
- Task
- Speech synthesis
- Parameters
- 40M
- Languages
- Chinese, English
- Highlights
- 15 built-in voices
- License
- Apache-2.0
Building it into a product? Talk to developer support.
The models are open source under Apache-2.0. For commercial support, write to developer@hojoai.com. To bring Hojo’s voice abilities to a device, connect it through Agentic OS.
Chat-Duplex and Omni are Hojo’s next generation of models.
Chat-Duplex
Full-duplex voice conversation you can interrupt at any time.
Omni
An omni-modal model for human–machine interaction.