AsrServe v1.0.4

v1.0.4 replaces the whole model stack on both the backend and the browser client. In our own tests against v1.0.3, recognition accuracy improved by about 20% and end-to-end efficiency by about 40%.

Breaking changes:

• Project rename: qwen3-asr is now AsrServe. The GitHub repository, Python package and Docker images (quantatrisk/asrserve) use the new name; old GitHub URLs redirect automatically.
• ASR model: Qwen3-ASR 1.7B/0.6B is replaced by Confucius4-R2T2. Offline and realtime share one R2T2 engine (vLLM on CUDA, vendored Rust on CPU); the model request field no longer switches models.
• Speaker diarization: CAM++ is replaced by Nemotron-3-Diarization (up to 8 speakers). Realtime streams now also carry per-utterance speaker labels.
• Removed models: FSMN VAD, the three CAM++ models and the Qwen3-ASR checkpoints are gone. Nemotron speech activity drives offline segmentation. ModelScope and FunASR are no longer dependencies; all models come from Hugging Face at pinned revisions.
• Removed API: the Alibaba Cloud compatible REST API (/stream/v1/asr*) is removed. Use the OpenAI-compatible /v1/audio/transcriptions and the native /v1/stream WebSocket.
• Deployment: images are published on Docker Hub as quantatrisk/asrserve:gpu (amd64), :cpu (amd64/arm64) and :ascend (arm64, Ascend 910B branch), plus 1.0.4-* version tags; build.sh builds from source. Compose files are compose.yml (GPU) and compose.cpu.yml (CPU) and no longer build. The only mount is ./models. docker-compose*.yml, deploy/prepare.sh and the model export option are removed. The old quantatrisk/qwen3-asr Docker Hub repository is no longer updated.
• Runtime: the CUDA image moves to CUDA 13.0 and vLLM 0.30; the service listens on port 17003.

Quick start:

docker compose up -d                        # GPU: quantatrisk/asrserve:gpu
docker compose -f compose.cpu.yml up -d     # CPU: quantatrisk/asrserve:cpu

macOS Apple Silicon runs natively without Docker; see the README.

🤖 Generated with Claude Code

v1.0.3

Changes

• Remove the voiceprint database, APIs, and sqlite-vec dependency.
• Make HF_HUB_OFFLINE the single offline deployment switch and use local model snapshots.
• Reduce default deployment configuration and align Compose, environment examples, and documentation.
• Increase internal cold-start readiness timeout to 600 seconds.
• Deprecate numeric weights in hotword context.

Qwen3-ASR v1.0.1

Changes

• Fix speaker-diarization long segment handling by splitting overlong CAM++ segments at low-energy boundaries before the ASR pipeline.
• Change the default offline ASR segment limit to 60 seconds and sync env examples/docs.
• Bump package and runtime version metadata to 1.0.1.

Validation

• uv run python -m compileall app/utils/speaker_diarizer.py app/core/config.py
• uvx pyright app/utils/speaker_diarizer.py app/core/config.py
• Speaker split smoke test on the 474s sample produced segments under 60 seconds.

Qwen3-ASR v1.0.0

Qwen3-ASR v1.0.0

English

Qwen3-ASR v1.0.0 is the first release after the project rename from funasr-api to qwen3-asr. The service is now centered on Qwen3-ASR, while keeping the existing FunASR/Paraformer realtime websocket compatibility path where it is still part of the runtime.

Highlights

• Renamed the project metadata, Docker Compose services, Docker image names, default logs, startup UI, benchmark reports, and core documentation to qwen3-asr.
• Updated GitHub Actions Docker publishing to push docker.io/quantatrisk/qwen3-asr:cpu-latest, docker.io/quantatrisk/qwen3-asr:gpu-latest, and docker.io/quantatrisk/qwen3-asr:latest.
• Set explicit Docker Compose project name qwen3-asr, so generated networks no longer depend on the local checkout directory name.
• Kept funasr==1.3.1, /ws/v1/asr/funasr, and FUNASR_* settings where they still represent real dependency, protocol, or runtime compatibility semantics.
• Preserved the current runtime split: CUDA uses official vLLM, CPU/macOS uses the vendored QwenASR Rust backend, and Paraformer remains the realtime-only websocket capability.

Docker Images

• docker.io/quantatrisk/qwen3-asr:latest
• docker.io/quantatrisk/qwen3-asr:gpu-latest
• docker.io/quantatrisk/qwen3-asr:cpu-latest

Upgrade Notes

• Update any deployment scripts, compose files, CI variables, and pull commands from funasr-api to qwen3-asr.
• If you rely on the legacy websocket compatibility endpoint, /ws/v1/asr/funasr is still intentionally available.
• If you configured FUNASR_* environment variables, keep them unchanged for this release.

中文

Qwen3-ASR v1.0.0 是项目从 funasr-api 改名为 qwen3-asr 后的首个正式发布版本。当前服务已经以 Qwen3-ASR 为核心，同时保留仍具备真实运行语义的 FunASR/Paraformer 实时 WebSocket 兼容路径。

主要变化

• 已将项目元数据、Docker Compose 服务、Docker 镜像名、默认日志、启动界面、benchmark 报告和核心文档统一改名为 qwen3-asr。
• 已更新 GitHub Actions 镜像发布配置，后续会推送到 docker.io/quantatrisk/qwen3-asr:cpu-latest、docker.io/quantatrisk/qwen3-asr:gpu-latest 和 docker.io/quantatrisk/qwen3-asr:latest。
• 已显式设置 Docker Compose project name 为 qwen3-asr，生成的网络名不再依赖本地目录名。
• 保留 funasr==1.3.1、/ws/v1/asr/funasr 和 FUNASR_* 配置名，因为它们仍然分别对应真实依赖、兼容协议和运行时配置语义。
• 当前运行时拆分保持不变：CUDA 使用官方 vLLM，CPU/macOS 使用 vendored QwenASR Rust backend，Paraformer 继续作为仅实时流式能力存在。

Docker 镜像

• docker.io/quantatrisk/qwen3-asr:latest
• docker.io/quantatrisk/qwen3-asr:gpu-latest
• docker.io/quantatrisk/qwen3-asr:cpu-latest

升级提示

• 请将部署脚本、compose 文件、CI 变量和镜像拉取命令中的 funasr-api 更新为 qwen3-asr。
• 如果仍依赖旧的 WebSocket 兼容端点，/ws/v1/asr/funasr 本版本仍会保留。
• 如果已经配置了 FUNASR_* 环境变量，本版本不需要改名。