Speech-to-Text
Local speech-to-text for Go
OpenAI-compatible /v1/audio/transcriptions and /v1/audio/translations powered by Bucky (whisper.cpp). Transcribe files, translate to English, stream live audio, and run models from tiny to large-v3-turbo with Metal, CUDA, and Vulkan — without sending audio to the cloud.
git clone https://github.com/ardanlabs/kronk.gitcd kronkmake example-bucky✓ installed whisper.cpp libraries✓ downloaded ggml-tiny.bin- text : And so my fellow Americans…- whisper.cpp
- GGML Whisper models
- Silero VAD
- Metal
- CUDA
- Vulkan
- WAV
- MP3
- FLAC
- SRT
- WebVTT
- OpenAI audio API
- Kronk model server
- Go SDK
Why Bucky
Built for production-minded Go teams
Bucky brings local speech-to-text into the Kronk stack so Go teams can transcribe and translate audio on their own hardware — without CGo or a cloud STT vendor.
Talk to an EngineerStay local
Transcribe and translate audio on your machines. Keep sensitive speech data inside your network.
Familiar endpoints
OpenAI-compatible /v1/audio/transcriptions and /v1/audio/translations so existing clients and agents can plug in with less friction.
Hardware accelerated
Metal, CUDA, and Vulkan acceleration for whisper.cpp models from tiny through large-v3-turbo.
Go-native path
No CGo. Pure-Go WAV, MP3, and FLAC decoding, plus a 1-to-1 whisper.h-style API for in-process use.
What you can do
Speech workflows Bucky is built for
Use the Go SDK inside your process, or expose the same capabilities through Kronk’s OpenAI-compatible audio APIs and Browser UI Translator.
Step 01
File transcription
Upload or decode WAV, MP3, or FLAC and get a transcript back through the SDK or HTTP API.
Step 02
Translate to English
Use /v1/audio/translations (or SDK translate mode) to turn supported source speech into English.
Step 03
Live streaming
Feed audio blocks and consume partial and final transcript events as speech arrives.
Step 04
Multi-speaker channels
Transcribe separate audio channels as separate speakers for meetings and call recordings.
API surface
OpenAI-compatible, running locally
Bucky powers Kronk’s audio endpoints, so clients that already speak the OpenAI audio API can transcribe and translate on your own hardware.
Explore the Bucky ManualCapabilities
What you get
whisper.cpp powered
Production-minded bindings for local transcription with models sized for your hardware and workload.
Streaming & VAD
Live streaming transcription with voice-activity detection — including Silero VAD for cleaner speech boundaries.
Diarization-ready
Channel-separated diarization in the SDK so multi-speaker recordings can map to separate speakers.
Flexible outputs
Return plain text, JSON, verbose JSON, SRT, or WebVTT — including optional word-level timestamps.
Works with Kronk server
Use Bucky through the Kronk model server, Browser UI Translator, or the Go SDK inside your process.
No cloud required
Run inference where your audio already is — laptops, workstations, or private infrastructure.
Models
Size the model to the job
GGML Whisper models from the bundled catalog — pull what you need with the Kronk CLI.
tiny / tiny.en
75 MBFastest path for prototyping and short clips
base / base.en
142 MBStrong speed-to-quality balance for everyday use
small / small.en
466 MBBetter accuracy when latency still matters
medium / large-v3-turbo
1.5 GBHigher accuracy without always paying full large-model cost
large-v3
2.9 GBHighest accuracy multilingual transcription when quality comes first
Get started
Run Bucky with Kronk
Install Kronk, pull whisper.cpp libraries and a model, then run the Bucky example — or hit the transcriptions API once the model server is up.
brew install ardanlabs/kronk/kronk# or: go install github.com/ardanlabs/kronk/cmd/kronk@latestkronk bucky libskronk bucky model pull tinygit clone https://github.com/ardanlabs/kronk.gitcd kronkmake example-buckyPlatforms
Where Bucky runs
Compatible whisper.cpp library bundles are downloaded per host. Use the CLI as the source of truth for supported combinations on your install.
macOS
- CPU
- amd64 · arm64
- GPU
- Metal (Apple Silicon)
Linux
- CPU
- amd64 · arm64
- GPU
- CUDA Vulkan
Windows
- CPU
- amd64
- GPU
- CUDA
Build AI systems you can own
Talk with Ardan Labs about local inference, private AI infrastructure, and production systems beyond API prototypes.