OpenASR Copyright 2026 OpenASR contributors This product includes software developed as part of the OpenASR project, licensed under the Apache License, Version 2.0 (see the LICENSE file). ClipFlow modifications: - Rebuilt OpenASR workspace version 0.1.23 from commit 7a5b3bdb59b50ebe2d83d870d636db5d0c064216 with an iOS 17 minimum deployment target. - Added one C FFI entry point that invokes OpenASR's existing idle native runtime-cache unload API after transcription. - Added an independent, reference-counted streaming cancellation-control handle and connected it to OpenASR's transcription control during session open, feed, and finish. This exposes the pinned upstream CPU ggml abort-callback support without concurrently accessing or freeing the mutable streaming-session handle. - Changed openasr-ggml's graph-allocation failure path to return an error when buffer reservation fails instead of continuing with a null buffer. Successful inference results and model weights are unchanged. - Reused prepared Whisper metadata, tokenizer, and materialized weights across VAD-delimited utterances in one streaming session, while leaving persistent shape-specific decoder-graph reuse disabled for that path. =========================================================================== Complete native dependency license materials =========================================================================== The complete versioned license materials for the native OpenASR artifact are published alongside this notice: openasr-ggml MIT license: https://www.clipflowapp.com/legal/openasr/v0.1.23-7a5b3bd/OpenASR-ggml-LICENSE.txt License archive for the exact Rust dependency closure: https://www.clipflowapp.com/legal/openasr/v0.1.23-7a5b3bd/OpenASR-ThirdParty-Licenses.txt Rust standard-library and compiler-builtins copyright notices: https://www.clipflowapp.com/legal/openasr/v0.1.23-7a5b3bd/Rust-Standard-Library-COPYRIGHT.html =========================================================================== Optional downloadable speech-recognition models =========================================================================== The following model packs are not bundled with ClipFlow. A user may download one or more of them directly through OpenASR's model distribution services. OpenASR converts and quantizes the upstream weights into `.oasr` packages and preserves the applicable upstream license and notice in each package. Whisper Tiny Q8 Upstream: openai/whisper-tiny OpenASR pack: OpenASR/whisper-tiny License: Apache-2.0 Whisper Base Q8 Upstream: openai/whisper-base OpenASR pack: OpenASR/whisper-base License: Apache-2.0 X-ASR Chinese-English Q8 Upstream: GilgameshWind/X-ASR-zh-en OpenASR pack: OpenASR/xasr-zh-en License: Apache-2.0 Whisper Small Q8 Upstream: openai/whisper-small OpenASR pack: OpenASR/whisper-small License: Apache-2.0 Whisper Large v3 Turbo Q4 Upstream: openai/whisper-large-v3-turbo OpenASR pack: OpenASR/whisper-large-v3-turbo License: MIT Copyright (c) 2022 OpenAI =========================================================================== Third-party components =========================================================================== openasr-ggml Path: crates/openasr-core/third_party/openasr-ggml License: MIT (see the published OpenASR-ggml-LICENSE.txt companion file) OpenASR's inference backend is a thin Rust layer built directly on ggml, a fork maintained in the openasr-ggml repository. ggml originates from the llama.cpp / whisper.cpp projects. Copyright (c) 2023-2024 The ggml authors Copyright (c) 2023-2024 Georgi Gerganov and contributors We gratefully acknowledge Georgi Gerganov and the ggml / llama.cpp / whisper.cpp communities. Their work is the computational core that makes this project possible. Improvements to the fork are contributed back upstream where applicable. FireRedVAD Stream-VAD Path: crates/openasr-core/src/diarize/vad/assets/firered_stream_vad_16k.safetensors License: Apache-2.0 OpenASR's sole neural Voice Activity Detection model: a causal (no lookahead) FSMN DetectModel (16 kHz), backing realtime endpointing, long-form speech slicing, and diarization. The vendored weights are the upstream Stream-VAD/model.pth.tar checkpoint and Stream-VAD/cmvn.ark CMVN stats re-serialized to safetensors; OpenASR reimplements the forward pass in pure Rust. Source: https://huggingface.co/FireRedTeam/FireRedVAD and https://github.com/FireRedTeam/FireRedASR2S (fireredvad package). Copyright 2026 Xiaohongshu We gratefully acknowledge the FireRedTeam for releasing a compact VAD model under a permissive license. pyannote community-1 PLDA/LDA parameters Path: crates/openasr-core/src/diarize/vbx/assets/community1_plda_f32.bin License: Creative Commons Attribution 4.0 International (CC BY 4.0) License URI: https://creativecommons.org/licenses/by/4.0/ Source: https://huggingface.co/pyannote-community/speaker-diarization-community-1/commit/8a527374977391da736e0daaef26855d949d9685 Attribution: pyannote Community-1 contributors. The upstream material credits the PLDA model to the BUT Speech@FIT group. The offline VBx-style diarization refinement includes PLDA/LDA numeric parameters from the public Hugging Face bundle at the revision linked above. Source files: https://huggingface.co/pyannote-community/speaker-diarization-community-1/blob/8a527374977391da736e0daaef26855d949d9685/plda/plda.npz https://huggingface.co/pyannote-community/speaker-diarization-community-1/blob/8a527374977391da736e0daaef26855d949d9685/plda/xvec_transform.npz Changes made: OpenASR converted the NPZ numeric arrays into one compact f32 runtime asset and implements the resegmentation logic independently in Rust. ClipFlow redistributes that OpenASR asset unchanged. No BUT VBx source code is vendored or copied. This attribution does not imply endorsement by pyannote, its contributors, or BUT Speech@FIT. Hy-MT2 1.8B translation model packs (staged metadata only) Upstream GGUF repository: tencent/Hy-MT2-1.8B-GGUF Upstream GGUF revision: 1cd5208700acedef4ef93019b6cfc148b8522d45 Base model repository: tencent/Hy-MT2-1.8B Base model revision: 9a341cd1b679d3efd23b46e847b01745a71ed792 License: Apache-2.0 OpenASR does not vendor Hy-MT2 weights in this source tree. If OpenASR distributes Hy-MT2 .oasr packs, each pack must include the upstream LICENSE.txt and the OpenASR NOTICE.openasr.txt modification notice. The staged PR1 recipe repackages upstream GGUF weights for local .oasr execution and does not modify model weights. JFK inaugural address audio (jfk.wav) Path: fixtures/jfk.wav License: public domain (US Government work); distributed via the MIT-licensed llama.cpp / whisper.cpp repositories The committed test/demo audio clip is `samples/jfk.wav` from the ggml-org llama.cpp / whisper.cpp projects: a public-domain excerpt of John F. Kennedy's 1961 inaugural address. OpenASR uses it as the single fixture for examples, the benchmark clip, and the test suite.