Features

ASRbox turns the small task of "transcribe this file" into a complete workflow.

Local transcription & batch tasks

Drop audio or video into the window and start. ASRbox preflights everything first: audio track presence, format support, duration and chunking — problems are explained before any work begins. Multiple files can be queued at once; tasks can be paused, resumed and retried.

15 local models: Whisper, Faster Whisper (word-level timestamps), MLX Whisper, SenseVoice, Qwen3-ASR, MOSS-Transcribe-Diarize (speaker diarization, 50+ languages)
Model sizes from 145 MB to 6.2 GB — pick your accuracy/speed trade-off
Resumable downloads: pause, resume and retry model downloads

ASRbox — Local transcription & batch tasks

GPU acceleration

No more waiting on long audio. With the optional NVIDIA CUDA kit on Windows, local transcription runs up to ~10× faster; on macOS the Apple GPU (Metal / MPS) is used automatically, and Apple Silicon can also pick the MLX-optimized models.

The CUDA kit is downloaded on demand — never installed if unused
Acceleration status is visible in Settings

ASRbox — GPU acceleration

Subtitle editing & version history

A transcript is never final until you say so. Fix text line by line on the timeline, batch-fix recurring mistakes with search & replace, and click any line to play from that exact spot. Every transcription, re-transcription, manual edit and AI pass creates a new immutable version — roll back to any step at any time.

Waveform audio player synced to each segment
Full-text search and batch replace
Complete, restorable version history

ASRbox — Subtitle editing & version history

AI subtitle proofreading

Connect any OpenAI-compatible LLM (presets for MiniMax, Kimi, DeepSeek, Qwen, GLM and Ollama) and let it read through your subtitles for typos, omissions and recognition errors. The AI only suggests: you review each change individually, and only accepted changes become a new version. The original transcript is always preserved.

Accept or ignore line by line — always in control
Only subtitle text with limited neighboring context is sent — never audio, file names or paths

ASRbox — AI subtitle proofreading

AI translation & bilingual export

Translate subtitles into any of 14 preset languages, or define your own target. Translations are separate versions too, and can be exported as bilingual subtitle files for learning and sharing.

14 preset languages plus custom targets
Bilingual subtitle export

ASRbox — AI translation & bilingual export

AI chat

Two ways to use it: unbound, it answers questions about using ASRbox itself (with a built-in offline knowledge base); bound to a transcript, you can ask about the content — "summarize this interview", "where do they mention budget" — with timestamped answers you can jump to.

Streamed replies, sessions stored locally
Quotes carry timestamps for quick verification

ASRbox — AI chat

Six export formats

TXT, SRT, VTT, ASS, JSON and Markdown. Into editing software, onto players, into notes, or piped into other tools — one of them always fits.

ASRbox — Six export formats

Online ASR & cloud LLMs

Beyond local models you can use online services: Bcut (free), Aliyun ASR, any OpenAI-compatible endpoint, or a self-hosted HTTP service. Whether to use them — and which one — is entirely your choice; the privacy notes list exactly what each scenario sends.

Online services are optional; everything is local by default
API keys stay on your machine

ASRbox — Online ASR & cloud LLMs

Model & storage management

Every model is labeled with size, language support and device compatibility: how much disk is left and whether a model will run on your machine is known before you download. Remove unused models anytime to reclaim space.

ASRbox — Model & storage management