What is the best local speech-to-text stack for my hardware?

As of October 2026: on a Mac, whisper.cpp (Metal and Core ML) for a command-line tool, WhisperKit for a Swift app, or parakeet-mlx when your languages are covered by NVIDIA's Parakeet models. On an NVIDIA GPU, faster-whisper for plain transcription, WhisperX when you need word timestamps and speaker labels, or NVIDIA NeMo for Parakeet and Canary models. On a CPU-only server, whisper.cpp with a quantized model or faster-whisper with int8.

whisper.cpp vs faster-whisper: which is faster?

It depends on the hardware. whisper.cpp is a C/C++ port with no Python dependency; on Apple silicon it runs fully on the GPU through Metal and can use Core ML, and it supports CUDA, Vulkan and integer quantization elsewhere. faster-whisper reimplements Whisper on the CTranslate2 engine and is up to 4 times faster than OpenAI's reference implementation at the same accuracy, with 8-bit quantization on CPU and GPU. On NVIDIA GPUs and Linux servers faster-whisper is usually the simpler fast choice; on Macs, whisper.cpp.

How do I get speaker labels (diarization) locally?

Use WhisperX on a GPU: it batches Whisper for up to 70x real-time transcription (with large-v2, per the project), aligns word-level timestamps with wav2vec2 and adds speaker diarization with pyannote. On Apple platforms, Argmax's open-source Swift package pairs WhisperKit with SpeakerKit for diarization.

Do I need a GPU for local transcription?

No, but it changes which model you can use comfortably. On a CPU-only machine, quantized Whisper small or large-v3-turbo through whisper.cpp, or faster-whisper with int8, handles recorded audio well. Real-time streaming of large models on CPU is hard; use a small streaming model such as Moonshine for live captions.

Quick Answer

Local Speech-to-Text Stack by Hardware: 2026 Setup Guide

Published: October 19, 2026

The short answer

Pick the stack by the machine first, then the model. As of October 19, 2026:

Apple silicon Mac: whisper.cpp (Metal + Core ML) for scripts and the command line; WhisperKit inside a Swift app; parakeet-mlx for NVIDIA’s Parakeet models in your languages.
NVIDIA GPU (Linux or Windows): faster-whisper for transcription; WhisperX for word timestamps and speaker labels; NVIDIA NeMo to run Parakeet and Canary.
CPU-only server or old laptop: whisper.cpp with a quantized model, or faster-whisper with int8.

The model ranking — which model is most accurate — is a separate question, answered in best local speech-to-text models ranked. This page is about the software that runs them.

The stacks compared

Stack	Runs on	Models	Strength	Latest release (GitHub)	Licence
whisper.cpp	Mac (Metal, Core ML), CUDA, Vulkan, CPU	Whisper family (GGML, quantized)	No Python, tiny footprint, Apple silicon first-class	v1.9.5 (Oct 6, 2026)	MIT
faster-whisper	NVIDIA GPU, CPU	Whisper family (CTranslate2)	Up to 4x faster than reference Whisper; int8	v1.2.1 (repo active Oct 2026)	MIT
WhisperX	NVIDIA GPU	Whisper + wav2vec2 alignment + pyannote	Word timestamps, speaker diarization, batched 70x real-time	v3.8.6 (May 25, 2026)	BSD-2
WhisperKit (Argmax OSS Swift)	iOS, macOS 14+	Whisper (Core ML); SpeakerKit, TTSKit in same package	Native Swift apps, streaming, on-device	v1.1.1 (Oct 8, 2026)	MIT
parakeet-mlx	Apple silicon (MLX)	NVIDIA Parakeet	Parakeet accuracy and speed on a Mac	active (Oct 2026)	Apache-2.0
NVIDIA NeMo Speech	NVIDIA GPU	Parakeet, Canary	Reference runtime for NVIDIA’s ASR models	v3.0.0 (Aug 7, 2026)	Apache-2.0

Setups that work

Mac, command line, any language. Build whisper.cpp, download large-v3-turbo in GGML format (quantize it for less memory) and run it; on Apple silicon inference runs fully on the GPU via Metal. Add the Core ML encoder for a further speed-up on the Neural Engine.

Mac app or iPhone. Add Argmax’s Swift package and use WhisperKit; its README recommends the compressed large-v3 (626 MB) checkpoint across iOS and macOS, and the same package includes SpeakerKit for diarization. For long recordings, its incremental mode streams audio from disk in bounded memory.

Mac, European languages, best accuracy. uv tool install parakeet-mlx and run Parakeet; Moondream’s Parakeet Ultra variant leads Apple-silicon dictation accuracy for 25 European languages.

GPU server, meeting recordings with speakers. WhisperX: batched transcription, wav2vec2-aligned word timestamps, pyannote diarization. You need a Hugging Face token for the pyannote models.

GPU server, plain high-volume transcription. faster-whisper with large-v3-turbo in float16, or int8 to fit more workers per GPU. For English-only maximum accuracy, run NVIDIA’s Canary-Qwen 2.5B through NeMo instead.

CPU-only. whisper.cpp or faster-whisper int8 with small or large-v3-turbo for batch jobs. For live captions on weak hardware, a streaming model such as Moonshine beats chunked Whisper.

How to choose in 30 seconds

Need speaker labels? WhisperX (GPU) or WhisperKit + SpeakerKit (Apple).
Shipping an iOS or macOS app? WhisperKit.
Language outside Parakeet’s 25 European languages? A Whisper-based stack — whisper.cpp or faster-whisper.
No Python allowed in production? whisper.cpp.
Live captions? A natively streaming model — see the streaming picks in the model ranking.

Benchmark on your own audio: accents, crosstalk and phone-quality recordings move word error rates far more than the stack does. The Open ASR Leaderboard scores models on one harness. If local is too much work, compare speech-to-text APIs, and for private meeting notes see local AI meeting note takers.

Last verified: October 19, 2026. Versions are each project’s latest GitHub release on that date; speed claims are the projects’ own.

The short answer

The stacks compared

Setups that work

How to choose in 30 seconds

Sources