JKinco Listen · 筑听（开源本地版）

Local model guide

Speech recognition (ASR)

Default: FunASR paraformer-zh + fsmn-vad + ct-punc.

Strong quality on Mandarin meeting audio, and it runs on CPU.
Downloaded automatically from ModelScope on first run (~1–2 GB).
Point JKINCO_ASR_MODEL_DIR at an existing model directory to skip the download.

Language model (LLM)

Default: Ollama qwen2.5:7b-instruct.

Hardware	Recommended model	What to expect
8 GB RAM	`qwen2.5:3b` / `qwen3:4b`	Usable, noticeably slow generation
16 GB RAM	`qwen2.5:7b-instruct`	The recommended default
32 GB RAM / 8 GB+ VRAM	`qwen2.5:14b` / `qwen3:8b`	Clearly better output quality
64 GB RAM / 24 GB+ VRAM	`qwen2.5:32b`	Approaching commercial quality

Switching models:

OLLAMA_MODEL=qwen2.5:14b docker compose up -d
# or, in manual mode
ollama pull qwen2.5:14b

Any OpenAI-compatible endpoint works — vLLM, llama.cpp server, LM Studio. Point .env at it:

LLM_BASE_URL=http://127.0.0.1:8000/v1/chat/completions
LLM_MODEL_NAME=your-model-name

Getting better output

Put proper nouns — contract sections, client organizations, recurring participant names — into JKINCO_ASR_EXTRA_TERMS, separated by 、. These are the words a general-purpose ASR model has never seen, and they are also the words whose errors are most visible in a finished document.
For meetings that matter, use a 14B model or larger. The gap between 7B and 14B shows up mostly in whether action items keep their owners attached.
Human review is the final source of truth. The system keeps the reviewed version alongside the original transcript so you can always see what changed.

A note on quantization

Ollama’s default quantization (Q4_K_M) is a reasonable trade-off for this workload. Minutes generation is a summarization task rather than a reasoning-heavy one, so the quality loss from 4-bit quantization is smaller here than it would be for code or math. If you have the memory headroom, moving up a model size buys more than moving up a quantization level.

This site is open source. Improve this page.