inference/research
verifiedhigh · 0.90updated 2026-10-04

whisper.cpp

cuda, rocm, hip, metal, vulkan, cpu-avx512, cpu-avx2, opencl, xnnpack, npu

Fields

id
engines/whisper-cpp
Repo
ggml-org/whisper.cpp
Languages
CC++
License
MIT
First release
2023
Design goal
high-performancecross-platformminimal-depsno-Python
Scheduling
none
Backends
cudarocmhipmetalvulkancpu-avx512cpu-avx2openclxnnpacknpu
Formats
gguf
Notable features
Plain C/C++ implementation with no dependencies beyond ggmlApple Silicon first-class citizen with Metal and Core ML (ANE) supportInteger quantization (Q5_0, Q5_1, Q8_0, etc.) for reduced memory and faster inferenceZero memory allocations at runtimeVulkan support for cross-vendor GPU coverage (AMD, Intel, NVIDIA)WebAssembly build for browser-based inferenceC-style API for easy integrationVoice Activity Detection (VAD) via Silero VADSupports all Whisper variants: tiny, base, small, medium, large-v1/v2/v3, large-v3-turboAlso supports NVIDIA Parakeet TDT models (src/parakeet.cpp)
Hardware caveats
NVIDIA: requires CUDA 11+ for GPU support; compute capability 5.0+AMD: ROCm/HIP support requires ROCm 5.6+; Vulkan is the more portable AMD pathCore ML (ANE) encoder support is Apple Silicon only and requires macOS 14+OpenVINO support is x86/Intel onlyQuantization can reduce accuracy; Q5_0 is the default for a reason
Best for
Cross-platform Whisper inference from a single binaryEdge and mobile deployment (iOS, Android, WASM)Scenarios where Python/PyTorch is not available or desiredRapid prototyping and benchmarking
Avoid for
Production multi-tenant serving with continuous batchingScenarios requiring the latest Whisper features or fine-tuned modelsNon-Whisper architectures (use vLLM or SGLang instead)
Docs
https://github.com/ggml-org/whisper.cpp
notes
whisper.cpp is a from-scratch C/C++ port of OpenAI's Whisper using the ggml tensor library. It is NOT a wrapper around PyTorch. The project is actively maintained with regular releases. It supports all Whisper variants and has recently added Parakeet TDT support. The Vulkan backend provides the widest GPU coverage. Core ML support on Apple Silicon can provide >3x speedup over CPU-only encoder inference. ||| MAINTENANCE (dated fact, GitHub API, 2026-10-04): last commit 2026-10-02 (0.1 months before this record), latest release v1.9.4 published 2026-09-11, 1043 contributors (distinct committers including anonymous), open_issues_count 350 (issues PLUS pull requests - see [[ev-github-open-issues-count-includes-pull-requests]]), not archived, 54127 stars. See [[ev-maintenance-status-is-four-dated-facts]] for how to read this fact set. ||| DEPLOYMENT: NO EVIDENCE OF PRODUCTION DEPLOYMENT FOUND as of 2026-10-04 - repositories, READMEs, docs and issue trackers were searched. See [[ev-production-evidence-is-scarce-and-nearly-always-single-site]] for the grading scheme; note that for an artifact-evaluation or paper-code release, non-deployment is documented intent rather than a negative finding.

Notes

whisper.cpp is a from-scratch C/C++ port of OpenAI's Whisper using the ggml tensor library. It is NOT a wrapper around PyTorch. The project is actively maintained with regular releases. It supports all Whisper variants and has recently added Parakeet TDT support. The Vulkan backend provides the widest GPU coverage. Core ML support on Apple Silicon can provide >3x speedup over CPU-only encoder inference. ||| MAINTENANCE (dated fact, GitHub API, 2026-10-04): last commit 2026-10-02 (0.1 months before this record), latest release v1.9.4 published 2026-09-11, 1043 contributors (distinct committers including anonymous), open_issues_count 350 (issues PLUS pull requests - see [[ev-github-open-issues-count-includes-pull-requests]]), not archived, 54127 stars. See [[ev-maintenance-status-is-four-dated-facts]] for how to read this fact set. ||| DEPLOYMENT: NO EVIDENCE OF PRODUCTION DEPLOYMENT FOUND as of 2026-10-04 - repositories, READMEs, docs and issue trackers were searched. See [[ev-production-evidence-is-scarce-and-nearly-always-single-site]] for the grading scheme; note that for an artifact-evaluation or paper-code release, non-deployment is documented intent rather than a negative finding.

Cross-references

This record references no other records — it is a leaf. It is cited by 1.