The speech-to-text zookeeper: one ggml runtime, 16 model families

What it does transcribe.cpp is a C++ inference library that loads speech-to-text models as GGUF files and runs them on the ggml runtime. It covers 16 model families—from Whisper and Parakeet to Moonshine, Canary, and Voxtral—across Metal, Vulkan, CUDA, and a tinyBLAS-accelerated CPU path. The project publishes pre-built GGUFs for every supported variant and wraps the core C API in bindings for Python, TypeScript, Rust, and Swift.

The interesting bit Most STT tooling treats Whisper as the default; transcribe.cpp treats it as merely one entry in a much larger catalog. The maintainers numerically verify and WER-test every ported model against its reference implementation, so the breadth does not come at the cost of correctness. It is essentially a cross-platform consolidation layer for an otherwise fragmented ecosystem of audio models.

Key highlights

• Supports 16 model families and 60+ variants, including streaming and batch modes.
• GPU backends cover Metal, Vulkan, and CUDA; CPU inference uses tinyBLAS kernels.
• Every published model is numerically verified and WER-tested against upstream references.
• Ships with a quantization tool and pre-built GGUFs hosted on Hugging Face.
• Single-header C API with official bindings for Python, TypeScript, Rust, and Swift.

Caveats

• Input audio must be 16 kHz mono WAV; other formats require external conversion.
• Some unit tests require a real model file to run, so the test suite is not fully self-contained.
• The medasr medical-dictation model is gated, which may complicate access.

Verdict Worth a look if you need to evaluate or deploy multiple STT models without maintaining separate inference stacks for each. Skip it if you are committed to a single-model pipeline and do not need the extra portability or cross-family comparisons.