// AI TOOLS · TRANSCRIBE.CPP

Transcribe.cpp

Open-source C/C++ library that runs 16+ speech-to-text model families locally on your own GPU via the ggml runtime — accurate, offline transcription with no per-minute API bill.

KompozyTurn one idea into a week of content — across every platform, published for you.
Get Started →

Last verified · 2026-07-19 · by Moe Ameen

What Transcribe.cpp is

Transcribe.cpp is an open-source C/C++ speech-to-text inference library built on the ggml runtime — the same runtime family behind llama.cpp — and the Mozilla.ai announcement frames it plainly as "llama.cpp for STT models." Instead of pinning you to a single model or a hosted API, it loads many transcription models from GGUF files and runs them locally on your own machine, with GPU acceleration through Metal (Apple Silicon), Vulkan (Linux/Windows), and CUDA (NVIDIA), plus a tinyBLAS-accelerated CPU path for hardware without a supported GPU. It was announced on the Mozilla AI blog on June 30, 2026, and is MIT-licensed.

The breadth is the point. Transcribe.cpp supports 16 ASR model families spanning 60-plus variants — including Whisper (12 variants, tiny through large-v3-turbo plus the English-only siblings), NVIDIA Parakeet and Canary, Moonshine, Qwen3-ASR, Cohere Transcribe, SenseVoice, GigaAM, and Voxtral, among others — and it handles both streaming and batch transcription. Because it reads GGUF, it's compatible with existing whisper.cpp model files, so a Whisper setup carries over. It ships language bindings for Python, TypeScript/JavaScript, Rust, and Swift/ObjC, and a `transcribe-cli` for command-line use; input audio is 16 kHz mono WAV.

It was built by CJ Pais (cjpais), the maintainer of Handy — a free, open-source, fully offline desktop speech-to-text app — under Mozilla.ai's Builders in Residence program. The project grew directly out of the pain of shipping cross-platform local transcription: many strong ASR models are developed in isolation with incompatible runtimes, and Transcribe.cpp gives them one uniform, GPU-accelerated interface. Every model published under the project's `handy-computer` Hugging Face organization is numerically validated and WER-tested against its reference implementation, so a model's output on Transcribe.cpp matches what the original produces.

The honest framing: this is a local transcription engine for developers and self-hosters, not a content tool. It turns audio into accurate text — offline, on your own hardware, with no per-minute bill — and it stops there. It writes no post, cuts no clip, designs no carousel, holds no brand voice, and publishes to nothing. It gives you the words; the rest of the job is still yours.

What you can make with it

  • Fully offline, private transcripts of podcasts, interviews, webinars, and voice memos — audio never leaves your machine
  • A whole backlog transcribed at zero marginal cost, since inference runs on hardware you already own
  • Word-level timestamped output usable for building subtitle tracks and caption alignment
  • Model-matched transcription — pick Whisper, Parakeet, Voxtral, or another family per audio type without changing tools
  • Streaming (live) or batch transcription inside your own app via Python, JS/TS, Rust, or Swift bindings
  • A drop-in upgrade path from an existing whisper.cpp setup, since it reads the same GGUF model files

How Kompozy turns Transcribe.cpp output into content

The unusual advantage of Transcribe.cpp is that transcription becomes effectively free and unlimited — it runs on hardware you already own, so a five-year archive of podcast episodes, a folder of webinar recordings, or every voice memo on your phone can all be turned into text at zero marginal cost, privately and offline. That changes the bottleneck. The question stops being "can I afford to transcribe this?" and becomes "what do I do with a mountain of transcripts?" — and a transcript is raw material, not content. Nobody reads a wall of timestamped text. That's the exact handoff Kompozy is built for: point it at a recording and it turns one source into 25–35 finished, on-brand outputs. Clipped Shorts finds the moments worth posting inside your long video and burns in styled captions; from the transcript text, Kompozy drafts a Blog Article and an Email Newsletter, pulls the sharpest lines into Quote Graphics and a brand-exact Carousel through HyperFrames, and writes native Text Posts — every piece governed by your Persona Brief so a rambling recording reads as your brand voice instead of verbatim dictation.

Run the two together and a local ASR run becomes a publishing pipeline. Transcribe locally with Transcribe.cpp so sensitive client calls and unreleased material stay on your machine, then bring each recording into Kompozy to reframe it to 9:16, 1:1, and 16:9 and fan it across nine social platforms plus a blog and Mailchimp newsletter from one queue, with Autopilot and a per-post review pipeline. Kompozy also generates formats a transcription library physically cannot — Persona Shorts and HeyGen avatar video with a face-locked recurring identity, Persona Frames, and Marketing Shorts — so the same idea can also ship as a talking-head clip. Transcribe.cpp gives you accurate words for free on your own GPU; Kompozy turns a whole archive of those words into months of scheduled, on-brand content.

  1. Run your recording (or a backlog of them) through Transcribe.cpp locally to get accurate transcripts offline, at no per-minute cost.
  2. Bring each recording and its transcript into Kompozy as your source.
  3. Let Kompozy cut Clipped Shorts with burned-in captions from the long recording, and draft a blog, newsletter, and text posts from the transcript.
  4. Pull the best lines into a brand-exact Carousel and Quote Graphics, held to one voice by your Persona Brief.
  5. Schedule and publish the whole set across TikTok, Reels, Shorts, X, LinkedIn, and more from one queue with Autopilot.

Frequently asked questions

What is Transcribe.cpp?

Transcribe.cpp is an open-source, MIT-licensed C/C++ speech-to-text library built on the ggml runtime — "llama.cpp for STT models." It runs 16+ ASR model families (60+ variants, including Whisper, Parakeet, Canary, Moonshine, and Voxtral) locally from GGUF files with GPU acceleration via Metal, Vulkan, and CUDA. It was announced on the Mozilla AI blog on June 30, 2026, and built by CJ Pais, the maintainer of the Handy transcription app.

Is Transcribe.cpp free, and does it run offline?

Yes on both. It is MIT-licensed open source with no per-minute API fee, and inference runs entirely locally on your own hardware — Mac, Windows, or Linux — so audio never leaves your machine and works without an internet connection once you have the model file.

How is Transcribe.cpp different from whisper.cpp?

whisper.cpp runs OpenAI's Whisper model. Transcribe.cpp is broader: it runs 16 ASR model families (60+ variants) — Whisper is just one of them — through one uniform, GPU-accelerated interface. Because it reads GGUF, it is compatible with existing whisper.cpp model files, so a Whisper setup carries over while you gain access to models like Parakeet, Canary, and Voxtral.

Can Transcribe.cpp create or publish social media content?

No. It transcribes audio into text and nothing more — it writes no posts, cuts no clips, designs no carousels, and publishes nowhere. To turn a transcript into captioned shorts, carousels, a blog, a newsletter, and scheduled posts across platforms, you use a content engine like Kompozy.

How do I turn a Transcribe.cpp transcript into content?

Transcribe the recording locally with Transcribe.cpp, then bring the recording and transcript into Kompozy. Kompozy cuts captioned Clipped Shorts from the audio, drafts a blog and newsletter from the transcript, builds carousels and quote graphics from the best lines in your Persona Brief voice, and schedules and publishes them across nine platforms plus blog and email.

Related tools

  • Whisper on Cloudflare Workers AIOpenAI's open-source Whisper speech-to-text model, served on Cloudflare's edge with a free daily allowance and per-audio-minute pricing.
  • Apple SpeechAnalyzerApple's on-device speech-to-text framework — a new proprietary transcription model, introduced at WWDC 2025, that benchmarks against OpenAI's Whisper.
  • OpenAI Voice Models (2026 update)OpenAI's 2026 voice stack in the API — real-time speech-to-speech agents, live translation, streaming transcription, and steerable text-to-speech, all in one family.
  • Krisp AIOn-device AI noise cancellation and a bot-free meeting assistant for clean audio.

← All AI tools · Get started →