FunASR v1.3.28: Reliable Realtime ASR and Better Subtitles

July 24, 2026 · FunASR weekly release

FunASR industrial speech recognition visual

FunASR v1.3.28 is a focused reliability release for live transcription and subtitle generation. Realtime WebSocket sessions now keep the best continuous transcript when a final decode regresses, while SenseVoice subtitle cues preserve text and timestamp alignment.

The signed release includes the Python wheel, source distribution, and nine self-contained llama.cpp / GGUF runtimes on one download page.

What changed

Scenario	v1.3.28 behavior	Why it matters
VAD-locked final decode becomes a short prefix, repetition, or exception	The server keeps the clean continuous partial covering the utterance	Final text no longer loses the end of a live sentence or replaces it with a repeated phrase
Short audio tail followed by `STOP`	The pending samples use the same final decode path	Short endings are returned instead of being silently dropped
WebSocket disconnect or handler error	Connection exceptions are imported explicitly and unexpected errors reach the generic handler	Error handling is stable across supported `websockets` versions and import order
SenseVoice subtitle segmentation	Rich tags, punctuation, and word/BPE timestamps stay aligned	Chinese is not collapsed into one cue and English surface text is preserved

1. Upgrade the Python package

Upgrade both realtime and subtitle deployments to the exact hotfix release:

python -m pip install -U "funasr==1.3.28"
funasr --version

The package version, command-line version, and server banner all report 1.3.28. Existing model caches and API request formats remain compatible.

2. Realtime and subtitle paths

For live Fun-ASR-Nano sessions, use the packaged realtime entry point and the repository's vLLM guide. Client-driven COMMIT and server VAD modes share the corrected finalization logic.

funasr-realtime-server --help
# See docs/vllm_guide.md for model and device options

The exact merged source passed 118 focused regression tests. The realtime WebSocket file also passed all 60 tests independently, including STOP final decode and import-order-independent exception handling.

SenseVoice users receive the subtitle alignment fix through the same package upgrade; no output-schema migration is required.

3. Download the self-contained llama.cpp / GGUF runtime

runtime-llamacpp-v0.1.9 ships nine Linux x64/ARM64, macOS ARM64, and Windows x64 packages. Linux includes AVX2 and Vulkan options; Windows includes AVX2, Vulkan, and CUDA, including funasr-llamacpp-windows-x64-vulkan.zip. Each archive provides a self-contained llama-funasr-* binary entry point.

Choose the standard CPU build when you do not want to maintain Python.
Choose Vulkan or CUDA for desktop GPU acceleration.
Choose ARM64 for edge devices and Apple Silicon.

See the llama.cpp / GGUF page for models, commands, and hardware selection. Runtime assets and model versions are linked together so users do not have to hunt across releases.

What each repository owns

Repository	What to follow
FunASR	Toolkit, serving, OpenAI API, vLLM, llama.cpp, and deployment docs
Fun-ASR	Fun-ASR-Nano / MLT models, capabilities, benchmarks, and model integrations
SenseVoice	Multilingual ASR plus emotion, language, and audio-event recognition
FunClip	Speech-driven video understanding, retrieval, and editing

Run one real audio file through the path that fits your deployment. File a reproducible GitHub issue when something blocks you, and star the repository you actually use.

Open the v1.3.28 release and downloads