March 2026 performance summary for k2-fsa/sherpa-onnx. Delivered stability and feature work focused on WebAssembly TTS, cross-language bindings for Supertonic TTS, and a formal release, driving reliability and broader adoption across platforms. Key improvements included memory safety in WebAssembly TTS, API usability enhancements, and expanded offline TTS capabilities through multi-language bindings and model uploads.

February 2026 — sherpa-onnx (k2-fsa). This month focused on stabilizing builds, expanding cross-language TTS/ASR capabilities, and accelerating model deployment across platforms. Key accomplishments include: - Stabilized Windows CI x64 pipelines and fixed CI issues, improving reliability for nightly builds and releases. - Fixed MSVC CRT support on Windows arm64, enabling robust ONNX Runtime builds on that architecture. - Expanded PocketTTS across languages with CXX and Swift APIs, Node.js bindings, and asynchronous JS APIs for generation and creation, delivering faster integration pathways for real-time TTS. - Launched Moonshine v2 multi-language API bindings across 14 languages with merged decoder support, plus export pipelines and WebAssembly deployment for web demos, broadening adoption and deployment options. - Strengthened packaging, runtime stability, and release hygiene (RPATH fixes, Python wheels, Windows PDB publishing, and CI hygiene), ensuring smoother distribution and fewer runtime issues.

January 2026 (k2-fsa/sherpa-onnx) delivered a broad platform and integration expansion with a focus on runtime stability, packaging reliability, and customer-facing capabilities. Key platform upgrades include upgrading onnxruntime to v1.23.2 across Linux variants (aarch64, arm, x64 with NVIDIA GPU) as part of migrating from 1.17.1, and addressing packaging and wheel issues (Linux ARM wheels, APK uploads). Major feature work includes Nemotron streaming support with APK packaging and ONNX export (including quantization and metadata) and expanding FunASR Nano across languages, with Go, JavaScript (node-addon), Java, Kotlin, Pascal, C#, WebAssembly bindings, plus a Dart API. PocketTTS integration matured with C++ runtime and Python support, Java/Kotlin APIs, a comprehensive CI/infra push, and a text length limiter; JNI refactor to improve type-safety and callback handling; new OfflineTts capabilities (getNumSpeakers) and broader Windows runtime support (MSVC MD/MT CRT options and shared build fixes). Stability and performance were enhanced through CI/test updates for FunASR Nano (C/C++ API and Node.js tests), packaging hygiene (HuggingFace URL redirects, Go CGO_ENABLED for examples, and VAD ASR APK build parallelism), and targeted fixes (Linux ARM wheels, APK upload fixes, Windows tokenizer checks). Overall, these efforts improve deployment speed, cross-platform availability, API coverage, and product reliability, driving faster time-to-market and broader customer adoption.

December 2025 performance summary for k2-fsa/sherpa-onnx. The team delivered a broad set of features that extend model support, improve performance, and enhance hardware acceleration, while addressing stability and build reliability across platforms. Key outcomes include end-to-end streaming demo support, expanded cross-language APIs for Google MedASR, and native runtime integrations on Qualcomm NPU, backed by code refactors and performance optimizations. A formal release (v1.12.20) closed the month with stable packaging and documentation updates. Highlights by category: - Features and model onboarding: streaming ASR demo for Paraformer, support for Fun-ASR-Nano-2512, WASM spaces for MatchaTTS, and multiple model exports (GigaAM v3, ZipVoice, Paraformer, MedASR) to sherpa-onnx and QNN paths. These enable broader deployment scenarios across cloud, edge, and mobile. - Hardware acceleration and runtimes: C++ runtime for Paraformer with Qualcomm NPU via QNN and Android demo integration, enabling lower latency and higher efficiency on mobile devices. - Architecture, performance, and quality: refactor ZipVoice C++ code, add 2-D matrix transpose, integrate Eigen for performance, and stabilize numerical computations to prevent NaN in speaker embeddings. - MedASR and multi-language bindings: backend APIs for Google MedASR across C/C++, Python, and multi-language bindings (Swift, C#, Pascal, Go, Dart, JavaScript/WebAssembly, Kotlin/Java), enabling broad adoption and consistent interfaces across languages. - Release and packaging: Release v1.12.20 with packaging improvements and bug fixes across the stack. Business value: These efforts expand model coverage (including MedASR and Nano variants), lower runtime latency on edge devices through QNN and Eigen optimizations, stabilize core pipelines to reduce crash/retry scenarios, and provide consistent cross-language APIs to speed integration for customer teams. The December work sets a durable foundation for scalable deployment, faster time-to-value for customers, and improved developer productivity through refactors and clearer API boundaries.

November 2025 (k2-fsa/sherpa-onnx). This month focused on accelerating hardware-accelerated deployment paths, expanding multi-language ASR capabilities, and strengthening release quality and tooling. The team delivered a broad set of features across Ascend NPU, QNN, and RK NPU ecosystems, while stabilizing builds and improving memory safety in key voice components.

Month 2025-10: Delivered Linux Desktop Streaming ASR support in the k2-fsa/sherpa-onnx Flutter UI, enabling streaming Automatic Speech Recognition on Linux desktops. Improved build pipeline and CI workflows with Android SDK/NDK versioning, version management, and caching to boost stability and developer productivity. Fixed Flutter app build reliability, enhancing local and CI build consistency. This work expands platform reach, reduces time-to-market for features, and demonstrates strong cross-platform development, CI/CD optimization, and Flutter tooling skills.

September 2025 monthly summary: Delivered cross-repo improvements and new TTS capabilities, and fixed critical reliability issues. Highlights include bias-term aware data writing in the model writer to prevent unnecessary bias data writes; fixes to sherpa-onnx wheel builds on aarch64 with CUDA/RKNN, including CI/CD workflow updates; introduction of the SA_miro Arabic TTS model for the JO locale with updated export workflow and metadata; and expanded CI coverage with non-streaming TTS tests for Zipvoice Go API. These efforts improved build stability, broadened platform support, and accelerated safe deployment of new TTS models and APIs.

July 2025 monthly summary for ml-explore/mlx focusing on accuracy in documentation. Delivered targeted documentation correction for mx.dequantize to align the documentation comment with the actual implementation. No code changes or API changes, preserving existing behavior. This reduces user confusion and support overhead, and improves docs quality for downstream guidance and tutorials.