Hojo-ASR-V1 – Hojo’s Open-Source Automatic Speech Recognition Model

1. What Is Hojo-ASR-V1?

Hojo-ASR-V1 is an open-source automatic speech recognition model from startup Hojo. Its four-stage hybrid stack combines Whisper feature extraction, Qwen3-Omni audio encoding, Conformer adaptation, and Qwen3-4B language-model decoding. Word error rate (WER) reaches 1.74% on LibriSpeech Clean, 7.6% on GigaSpeech, and 7.02% on VoxPopuli—approaching top commercial models such as NVIDIA Canary and IBM Granite. Fully local deployment targets AI Agent workflows, voice input, and privacy-sensitive scenarios.

Technical positioning and domain: ASR, but architected as an LLM-driven semantic speech system—not acoustic matching alone. Qwen3-4B decoding adds semantic understanding and context correction for colloquial speech, mixed Chinese–English, and domain terminology while staying competitive on standard benchmarks.

Research background: Hojo’s team spans speech, LLMs, and systems. Pain points: Whisper-class models struggle with colloquial homophones and noise; end-to-end multimodal models (e.g., Qwen3-Omni) are strong semantically but costly and slow. Hojo-ASR-V1 balances Whisper acoustics + Qwen3-4B semantics via Conformer adaptation.

Core value: Semantic correction and local privacy. Homophones like “公式” vs “攻势” disambiguate via language priors; all audio stays on-device for meetings, medical, and finance compliance.

Technical characteristics: Encoder–adapter–LLM decoder; Conformer aligns dimensions and compresses ~1500 Whisper frames (30s) to ~300 frames, cutting LLM compute. Tuned for Chinese colloquialisms (轻声, 儿化, fillers).

2. Key Features

• High-accuracy transcription with semantics: Beyond acoustic mapping—contextual correction (e.g., “公式 meeting” → “攻势 meeting”). LibriSpeech Clean WER 1.74%.

• Chinese–English code-switching: Robust on mixed utterances like “configure the API endpoint token”—~30% WER reduction vs traditional models in mixed scenes.

• Noisy environments: LLM semantics compensate for corrupted acoustics; GigaSpeech WER 7.6%.

• Domain terminology: Medical, legal, tech, finance terms via Qwen3-4B knowledge—less post-editing.

• Fully offline: Local GPU/CPU; Apache-2.0 weights on Hugging Face and GitHub; commercial use allowed.

• Streaming: Real-time chunks with second-scale latency for Agents, live captions, and voice IME.

3. How to Use

1. Environment: Linux (Ubuntu 20.04+) recommended; Python 3.10+; GPU ≥8GB VRAM (RTX 3070+) preferred; CUDA 11.8+ / cuDNN 8.x; ~10GB disk.

2. Dependencies: pip install hojo_asr torch torchaudio transformers (match CUDA PyTorch wheels as needed).

3. Weights: https://huggingface.co/HojoAI/Hojo-ASR-V1 (~8GB); auto-download via Transformers or manual cache.

4. Load model:

   import hojo_asr
   model = hojo_asr.load_model("HojoAI/Hojo-ASR-V1", device="cuda")  # or "cpu"
   

5. Transcribe:

   result = model.transcribe("audio.wav")  # wav, mp3, flac, ...
   print(result["text"])
   

   Stream with model.transcribe_stream(). Optional post-polish via DeepSeek/GPT.

6. Config: beam_size (default 5), language (auto/zh/en), temperature (0.0 greedy). Long audio auto-segments; chunk_size=30 explicit if desired.

4. Pros and Cons

5. Comparison with Similar Tools

Selection advice: Privacy + accuracy (enterprise meetings, medical, finance)—Hojo-ASR-V1. Ultra-low latency (<500ms) assistants—Whisper small/medium. Cloud scale without data residency—Canary/Granite at licensing cost.

6. Editor’s Take

Hojo-ASR-V1 exemplifies “acoustic + semantic” ASR—not full Qwen3-Omni end-to-end, but Whisper acoustics + Qwen3-4B semantics via Conformer—a pragmatic split balancing cost and quality.

1.74% LibriSpeech is near commercial grade; semantic wins show most in messy real speech. Local deploy solves compliance cloud ASR cannot. ~8GB and ~1–2s latency limit edge and sub-100ms use.

Audience: Agent builders, enterprise IT (local minutes), voice app startups.

— Hardware and ecosystem deductions; architecture and accuracy highly compelling for privacy-conscious teams.

7. Use Cases

• Agent voice entry: Spoken commands → executable text in coding Agents and assistants.

• System-wide voice IME: OS-level dictation into browsers, ChatGPT, Notion, VS Code—local and terminology-aware.

• Meeting transcription: Offline minutes with timestamps; optional LLM summarization—no cloud audio upload.

• Podcast / video captions: Long-audio segmentation; SRT/VTT export; strong on zh/en mixed subtitles.

• Call center QA: Real-time dialogue transcription for compliance and analytics despite phone noise.

8. FAQ

Q: vs Whisper?
A: Semantic disambiguation, colloquial/mixed-language robustness, Chinese tuning—~30% mixed WER improvement vs Whisper in cited scenarios.

Q: Minimum hardware?
A: ≥8GB NVIDIA GPU ideal (~1–2s / 30s). CPU 16GB+ works (~5–10s); shrink beam_size to 1. No native AMD/MPS acceleration—CPU fallback only.

Q: Formats and max length?
A: WAV, MP3, FLAC, OGG; auto-chunks >30s; hours-long feasible but slow—compress/downsample if needed.

Q: Fine-tuning?
A: Official fine-tune tooling in progress; Apache-2.0 stack allows custom Hugging Face fine-tunes; domain lexicons/prompting as interim.

Q: Colloquial / mixed zh-en?
A: Core strength—fillers and code-switch handled well.

Q: Speed optimizations?
A: Lower beam_size, FP16, watch for future INT8/INT4 releases.

9. Project Links

• GitHub: https://github.com/HojoAI/Hojo-ASR
• Hugging Face: https://huggingface.co/HojoAI/Hojo-ASR-V1