Hojo-ASR-V1 - Hojo open-source automatic speech recognition model

What is Hojo-ASR-V1?

Hojo-ASR-V1 is an open-source automatic speech recognition model from the startup Hojo. It employs an architecture combining Whisper feature extraction, Qwen3-Omni audio encoding, Conformer adaptation, and Qwen3-4B language model decoding. The model achieves a word error rate of only 1.74% on the LibriSpeech Clean dataset, 7.6% on GigaSpeech, and 7.02% on VoxPopuli, demonstrating performance close to top-tier software. The model supports local deployment and is designed for agent workflows and voice input scenarios.

Main functions of Hojo-ASR-V1

• High-precision speech transcriptionIt converts audio to text in real time, supporting long sentences and complex semantic understanding.
• Chinese and English mixed recognitionIt exhibits strong robustness to colloquial expressions and scenarios involving a mix of Chinese and English.
• Noise environment adaptationWith the help of semantic judgment by language model, it still maintains a high recognition rate in noisy environments.
• Understanding technical termsUsing the semantic capabilities of Qwen3-4B, accurately identify proper nouns in fields such as medicine, law, and technology.
• Local offline operationIt supports fully local deployment and can complete voice recognition without an internet connection, ensuring data privacy.

Technical Principles of Hojo-ASR-V1

• Acoustic feature extractionThe audio is processed by the feature extractor of OpenAI Whisper, which converts the original waveform into a high-dimensional acoustic feature vector while preserving phoneme and spectral information.
• Audio semantic codingThe feature vector is input into the Qwen3-Omni audio encoder for deep semantic encoding, which transforms the sound information into a hidden state representation aligned with the text semantic space.
• Feature adaptation and compressionThe intermediate layer adopts the Conformer structure to perform feature adaptation and temporal compression between the encoder and the language model, balancing computational efficiency and information fidelity.
• Large Language Model DecodingThe features received and processed by the Qwen3-4B large language model are combined with semantic context to generate the final text, enabling the model to use language priors to correct errors caused by similar pronunciations or noise interference.

How to use Hojo-ASR-V1

• Download ModelAccess the HuggingFace repository HojoAI/Hojo-ASR-V1 Or GitHub HojoAI/Hojo-ASR Obtain weights and code.
• Install dependenciesConfigure the Python environment and install... hojo_asr and related dependency libraries.
• Loading Model:pass hojo.load_model() The interface loads the model onto the local GPU or CPU.
• Input audio: Pass in the audio file path or real-time audio stream, and call the transcription interface.
• Get ResultsThe model returns recognized text, which can be integrated with large models such as DeepSeek and GPT for secondary polishing and format optimization.

The core advantages of Hojo-ASR-V1

• Top-notch recognition accuracyThe LibriSpeech Clean WER is 1.74%, which is close to the level of top models on the leaderboards such as NVIDIA Canary and IBM Granite.
• Semantic error correctionUnlike traditional ASR which relies solely on acoustic matching, the Qwen3-4B language model can infer the correct words based on contextual semantics, significantly reducing homophone errors.
• Chinese scene optimization: It has been deeply optimized for details such as colloquial Chinese expressions, neutral tones, and retroflex endings, resulting in a better user experience than general multilingual models.
• Privacy and securityIt supports fully local deployment, and sensitive voice data does not need to be uploaded to the cloud, making it suitable for enterprise meetings and medical scenarios.

Hojo-ASR-V1 project address

• GitHub repositoryhttps://github.com/HojoAI/Hojo-ASR
• HuggingFace model libraryhttps://huggingface.co/HojoAI/Hojo-ASR-V1

Comparison of Hojo-ASR-V1 with similar competing products

Application scenarios of Hojo-ASR-V1

• AI Agent Voice EntryAs the "ears" of the intelligent agent, it converts voice commands into executable text, replacing the keyboard as the primary mode of interaction.
• System-level voice input methodTake over global input in the operating system and enable voice typing in any text box in browsers, ChatGPT, Claude, Notion, etc.
• Real-time meeting transcriptionLocal deployment enables offline meeting minutes generation, preventing the leakage of confidential business audio data.
• Podcasts and Video CaptionsQuickly convert long audio content into text transcripts, supporting subsequent large-scale model polishing and structured organization.
• Intelligent customer service and telephone quality inspectionIn call center scenarios, it can recognize customer voice in real time and extract key information and emotional keywords.