You can deploy this model and start earning money today!

Mininglamp-2718/
OctoASR-1.7B-Instruct-1.0-MLX-8bit

Name: OctoASR-1.7B-Instruct-1.0-MLX-8bit
Availability: InStock

safetensors

qwen3_asr

8-bit

Views

No views yet

OctoASR

MLX

中文 | English

Introduction

OctoASR is a local speech recognition service for vertical domains, deeply optimized for Apple Silicon via Cider. Purpose-built for Internet or IT office scenarios, it is closely adapted to high-frequency workplace use cases such as meeting notes, technical discussions, product reviews, and engineering dictation. Through targeted optimization on domain data, OctoASR accurately recognizes English terms, acronyms, and product names (e.g. Kubernetes, FastAPI, PRD, Code Review) as well as mixed Chinese-English speech, effectively addressing the term-misrecognition and code-switching segmentation issues common to general-purpose models — so transcripts come out clear, domain-aware, and accurate. The service runs fully locally, works out of the box, and keeps audio and transcript data on your machine.

Core capabilities:

🎯 Vertical-domain optimization — optimized on Internet or IT office data; accurate on English terms, acronyms, product names, and mixed Chinese-English speech.
🍎 Native Apple Silicon — MLX-based local inference on M-series chips, further optimized with our in-house acceleration framework Cider.
🔒 Fully local, privacy-first — audio and transcripts never leave your machine.
✂️ VAD segmentation — optional FSMN VAD splits long audio and transcribes segment by segment.
🧩 Pluggable engines — supports Fun-ASR-Nano, Qwen3-ASR and more base models, switchable with one command.
🏷️ @Mention replacement — auto-fix nicknames and transliterated names in transcripts via a visual page. See Mentions.
⚡ One-command start — install via brew install, then octoasr start.

Changelog · Models · Installation · Usage · API · License · Acknowledgments ·

Changelog

See the full release history on the Releases page.

2026-06-15 — Release the first ASR model built on Qwen3-ASR and tailored for internet office scenarios, supporting handwritten-style transcription output and accurate recognition of industry terminology.
2026-06-09 — Added @mention replacement with a visual management page (octoasr mentions) for editing nickname → canonical-name mappings; spoken "艾特" is normalized to @ before replacement. (v0.1.15 fixes packaging so the web page ships in the Homebrew build.)
2026-05-29 — Released the first ASR model for internet office scenarios, with written-style transcription output and accurate recognition of industry-specific terminology.
2026-05-26 — First release: FastAPI transcription service, FunASR-Nano engine, FSMN VAD, hotword extraction, session logging.

Models

OctoASR uses a pluggable engine design and supports several mainstream ASR base models. Switch with a single command: octoasr model use <name>.

Model	Base model	Quant	Size	Languages	Links
OctoASR-1.7B (default)	Qwen3-ASR	8bit	1.7 GB	ZH / EN	🤗 · 🤖 · 🌟
OctoASR-0.8B	Fun-ASR-Nano	8bit	0.8 GB	ZH / EN	🤗 · 🤖 · 🌟

The model is downloaded automatically from HuggingFace or ModelScope (China mirror); the source is chosen by network environment on first run.

Installation

Option 1: Homebrew (recommended)

bash

brew tap Mininglamp-AI/tap
brew install octoasr

# Start (first run auto-initializes + downloads the default model)
octoasr start
octoasr doctor   # environment check

Option 2: From source

bash

# 1. Dependency: ffmpeg (decodes non-WAV audio)
brew install ffmpeg

# 2. Clone + install
git clone https://github.com/Mininglamp-AI/OctoASR.git
cd octoasr
python3 -m venv .venv && source .venv/bin/activate
pip install -U pip
pip install -e .

# 3. Download the model
hf download Mininglamp-2718/OctoASR-1.7B-Instruct-1.0-MLX-8bit \
  --local-dir models/Mininglamp-2718/OctoASR-1.7B-Instruct-1.0-MLX-8bit

hf download Mininglamp-2718/fsmn-vad-mlx \
  --local-dir models/Mininglamp-2718/fsmn-vad-mlx

# Behind a China mirror:
# HF_ENDPOINT=https://hf-mirror.com hf download ...

# 4. Start the server
python3 server.py \
  --model-path models/Mininglamp-2718/OctoASR-1.7B-Instruct-1.0-MLX-8bit \
  --vad-model-path models/Mininglamp-2718/fsmn-vad-mlx \
  --host 0.0.0.0 --port 8787 --load-on-startup

Requirements: macOS (Apple Silicon) · Python 3.10+ · ffmpeg / ffprobe on PATH.

Usage

The example below shows audio translation: transcribe speech and translate it into Chinese.

CLI (recommended)

bash

# On first run, the service auto-initializes and downloads the default model
octoasr start

# Transcribe / translate an audio file
octoasr transcribe assets/BAC009S0764W0129.wav

Python API

python

from core.auto_model import AutoModel

model = AutoModel(
    model="models/Mininglamp-2718/OctoASR-1.7B-Instruct-1.0-MLX-8bit",
    vad_model="models/Mininglamp-2718/fsmn-vad-mlx",  # optional: auto-segment long audio
)

text = model.generate(
    "assets/BAC009S0764W0129.wav",
    task="translate",  # translation task
    target_language="zh",  # target language: Chinese
    merge_vad=True,
)
print(text)
# -> "甚至出现交易几乎停滞的情况"

HTTP API

bash

curl -X POST http://127.0.0.1:8787/v1/voice/transcribe \
  -F "audio=@assets/BAC009S0764W0129.wav" \
  -F "personal_context=## Terms\n- FastAPI\n- Kubernetes" \
  -F "mode=smart"

json

{
  "status": 200,
  "text": "transcribed text",
  "m": "octoasr",
  "engine": "mlx"
}

Full API fields, limits and auth are documented under API.

@Mention replacement

Auto-replace casual nicknames and transliterated names in transcripts with the canonical spelling you want (e.g. @小明 → @Xiaoming). Manage entries on a visual web page — no JSON editing required:

bash

octoasr start        # Start the service (if not running)
octoasr mentions     # Open the management page in your browser

📖 Full guide: Mentions

API

`POST /v1/voice/transcribe`

Transcribe a single uploaded audio file. Request type: multipart/form-data.

Field	Type	Required	Description
`audio`	file	yes	Audio file. Supported: `.wav` `.mp3` `.ogg` `.webm` `.m4a` `.flac`
`context_text`	string	no	Existing text for append/edit modes; last 5000 chars kept
`chat_context`	string	no	Chat context; last 20000 chars kept
`personal_context`	string	no	Personal correction / hotword context; last 10000 chars kept
`member_context`	string	no	Member context; last 5000 chars kept
`mode`	string	no	`smart` / `append_only` / `edit_only`, default `smart`

Limits: default max file 30 MiB, max duration 660 s; edit_only requires context_text.

`GET /v1/voice/config`

Returns current service limits and engine metadata.

curl http://127.0.0.1:8787/v1/voice/config

Authentication

Disabled by default. If started with --auth-token, requests must carry Authorization: Bearer <token>.

python3 server.py --model-path <path> --auth-token "$OCTOASR_TOKEN"

License

Released under the MIT License.

Acknowledgments

OctoASR would not be possible without these excellent open-source projects:

MLX & mlx-audio — Apple's machine-learning framework and audio toolkit, the foundation of OctoASR's local inference.
FunASR / FunAudioLLM — source of Fun-ASR-Nano and FSMN-VAD, providing strong Chinese speech recognition.
Qwen3 — the base model behind the Qwen3-ASR engine.
mlx-community — high-quality MLX quantized models.
ModelScope & Hugging Face — model hosting and distribution.
FastAPI — high-performance web framework.

Thanks to everyone contributing to the open-source speech recognition community.