Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarHabr AI→ original

Wavo in Rust Transcribes an Hour of Audio in 18 Seconds, 200x Faster Than Real Time

The author couldn't find a native dictation app, so with AI agents he wrote wavo, a speech recognition engine in Rust. On a MacBook Pro with the M5 chip, it uses Apple's Neural Engine to transcribe an hour of recording in roughly 18 seconds: 200× faster than real time. On a GPU, it reaches 160× against 107× for transcribe.cpp. The shaders were written once with wgpu, with no C++ and no per-GPU code. The takeaway: agents can already write shaders, but speed has to be measured on every piece of hardware.

AI-processed from Habr AI; edited by Hamidun News
Wavo in Rust Transcribes an Hour of Audio in 18 Seconds, 200x Faster Than Real Time
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Custom FFT176 lines instead of 55,000 in rustfft, rebuild time from 11.9 to 2.1 s
Whisper turbo on RTX 4060a short recording in 201 ms versus 226 ms for transcribe.cpp
Russian GigaAM v3330× real time versus 262× for transcribe.cpp

What does "200 times faster than real time" mean? An hour-long recording is transcribed in about 18 seconds (3,600 seconds divided by 200). That result came from a developer who only wanted a convenient dictation app but ended up writing his own speech recognition engine in Rust, wavo, with the help of AI agents. It runs on a MacBook Pro with the M5 chip and outperforms popular ready-made solutions. The story is useful to anyone building local AI tools: it shows where agents genuinely help and where you can't do without testing on real hardware.

How It Started

The author switched to a MacBook Air M5 and, at the same time, decided to replace heavy software with native alternatives: VS Code with Zed, Chrome with Helium. For dictation while working with agents, he used Handy, but it is built on Tauri, and he wanted a native app. The agent couldn't find a suitable alternative, so the developer set out to write his own. Meanwhile, he tried different interface options (GPUI Kit, egui, system controls) and ultimately wrote his own UI library: the rendering alone took about 3,000 lines of Rust, and the basic components another 6,000.

The core of the project is the speech recognition engine. The ready-made transcribe.cpp linked for about a minute on every rebuild, and C++ came along into the Rust project. Inside was the ggml library, which has separate code and build tools for each GPU backend (Metal, Vulkan, CUDA). The alternative was wgpu: shaders are written once in WGSL, run on any GPU, and build with a regular cargo build. The author tried running the GigaAM v3 model on it; the agent wrote the shaders easily, and the result was comparable to the original.

An important detail about the GGUF format: it stores only weights and metadata (architecture name, layer sizes, vocabulary). The computations themselves aren't in the file, unlike TensorFlow or Core ML, which describe the operation graph. So the computations have to be written separately for each model family: GigaAM, Parakeet, Whisper. The author had to work out how this all fits together: audio is split into chunks, converted into a spectrogram (which frequencies are present at each moment) via the fast Fourier transform, an encoder turns the spectrogram into vectors, and a decoder (CTC, RNN-T, or TDT) assembles the text from them.

Codex wrote the first version of Parakeet v2 support overnight, but the code was messy. The author threw it out and rewrote it with Claude Code. Then Parakeet v3 and Whisper were added as a reference for comparison.

Where the Speedups Came From

Before publishing, the author turned to build speed. The rustfft crate slowed down rebuilds, and when wavo was used as a library, every user would have paid that delay. He wrote his own FFT for the single task at hand: 176 lines of code versus 55,000 in rustfft, and rebuild time dropped from 11.9 to 2.1 seconds. The author praises rustfft itself: it is designed for every case, while he needed just one.

Tests on different hardware revealed a failure on Windows with an NVIDIA GPU. The cause was flash attention. Standard attention writes the table of comparisons between all vectors in an audio chunk to GPU memory and reads it back. Flash attention computes the same thing in small tiles in fast memory next to the cores and never stores the full table. On Macs, the agent had already written such shaders for Apple's matrix blocks, but NVIDIA's blocks are a different size, so the code took the standard path. After adding NVIDIA-specific shaders, wavo transcribes a short Whisper turbo recording in 201 ms versus 226 ms for transcribe.cpp (it was 637 ms before).

On the M5, another finding: each GPU core now includes a neural accelerator that multiplies matrices three times faster (0.60 ms versus 1.89 ms). But access to it is only available through Metal 4, and wgpu doesn't support it yet. The author got around the limitation with a separate shader on Metal, used only for the M5, and opened an issue in the wgpu project.

Numbers and Limitations

In a comparison on English with Parakeet v3 (MacBook Pro M5), wavo on the Neural Engine, Apple's dedicated block for neural networks, transcribes a phrase of 5 seconds in 0.12 s and an hour of recording at 200 times real time. FluidAudio on the Neural Engine reaches 175×, wavo on the GPU 160×, transcribe.cpp on the GPU 107×, parakeet.cpp 90×, parakeet-mlx 61×, and sherpa-onnx on the CPU 39×. On Russian with GigaAM v3, the gap is similar: wavo is 330 times faster than real time, transcribe.cpp 262 times.

The caveats matter. The Windows measurements were made on a single RTX 4060 GPU; the author did not test AMD Radeon at all and asks for results to be submitted as issues. The table comes from one Mac, and the project is a personal one, not an independent benchmark.

The practical takeaway: agents are already capable of writing shaders and an entire inference engine, provided a human understands the model architecture and verifies the results. But the most costly discoveries, such as the NVIDIA failure and the unused M5 blocks, surfaced only on real hardware. If you're building a local AI tool, budget time for testing on different GPUs.

Story

Key Questions Answered

How It Started?

The author switched to a MacBook Air M5 and, at the same time, decided to replace heavy software with native alternatives: VS Code with Zed, Chrome with Helium. For dictation while working with agents, he used Handy, but it is built on Tauri, and he wanted a native app. The agent couldn't find a suitable alternative, so the developer set out to write his own.

Where the Speedups Came From?

Before publishing, the author turned to build speed. The rustfft crate slowed down rebuilds, and when wavo was used as a library, every user would have paid that delay. He wrote his own FFT for the single task at hand: 176 lines of code versus 55,000 in rustfft, and rebuild time dropped from 11.9 to 2.1 seconds. The author praises rustfft itself: it is designed for every case, while he needed just one. Tests on different hardware revealed a failure on Windows with an NVIDIA GPU. The cause was flash attention.

What are the key details on Numbers and Limitations?

In a comparison on English with Parakeet v3 (MacBook Pro M5), wavo on the Neural Engine, Apple's dedicated block for neural networks, transcribes a phrase of 5 seconds in 0.12 s and an hour of recording at 200 times real time. FluidAudio on the Neural Engine reaches 175×, wavo on the GPU 160×, transcribe.cpp on the GPU 107×, parakeet.cpp 90×, parakeet-mlx 61×, and sherpa-onnx on the CPU 39×. On Russian with GigaAM v3, the gap is similar: wavo is 330 times faster than real time, transcribe.cpp 262 times.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…