Self-Hosted Transcription / STT
As of September 2026, AIDiveForge tracks 12 self-hosted transcription / stt. The top three by verified-data score are VoxRT Wake-Word, CosmoWhisper, and AirGapScribe. Curated self-hosted transcription / stt tracked by AIDiveForge. Listings are verified against each tool's live website and re-checked regularly.
Last updated September 16, 2026 · 12 tools
Ranked by AIDiveForge's verified-data score: data completeness, verification recency, community rating, and real visitor engagement. How we rank · No tool can pay for placement.
1. VoxRT Wake-Word
The SDK ships a Rust runtime under 1 MB with wake-word models around 100 KB, so it fits on mobile and IoT targets without gutting your memory budget. Audio stays on the device — the vendor states models are encrypted at rest and the system works offline by default, which means GDPR and HIPAA conversations get simpler, not harder. The published models are free for commercial use; custom models trained to your phrase, accent profile, or domain vocabulary are a paid engagement. iOS and Android are available in v1; Windows, WebAssembly, microcontrollers, automotive, and wearables are listed as v2, meaning shipping on those targets today is not an option. Teams that need a language other than English are also waiting — multilingual support is post-v1 on the roadmap.
PaidSelf-hostedVerified Sep 9, 2026
2. CosmoWhisper
Built in C#/.NET with the Windows App SDK, it idles under 90MB RAM and delivers sub-500ms transcription across Slack, Word, Outlook, Notion, and VS Code without an Electron runtime weighing it down. The local offline mode — called Race Mode — runs a Whisper server entirely on-device, so audio never leaves the machine, which is the architecture medical and legal teams need for HIPAA compliance. Smart Commands let you highlight text and say 'Fix grammar' or 'Summarize' for inline edits. The free tier caps at 60 minutes per month, which is enough for evaluation but not for a full workday. Teams that need cross-platform coverage — a Windows desk paired with a MacBook — hit a hard wall immediately, as the vendor states Windows 10/11 is the only supported OS.
Paid$12 / moSelf-hostedVerified Sep 9, 2026
3. AirGapScribe
The core workflow is a Windows tray app: record from microphone, system audio, or both; transcription runs on-device using Whisper models; export as .txt. The free demo caps sessions at ten minutes and watermarks exports — enough to verify the local pipeline, not enough for production use. The paid tier removes those caps, adds an on-device AI assistant that queries across saved transcripts, generates structured deliverables, and accepts .txt, .md, .pdf, or .docx files as context. NVIDIA GPU acceleration is supported and cuts generation time significantly, but CPU-only machines still work. There is no API, so AirGapScribe does not plug into existing pipelines — you pull deliverables out manually.
Paid5.99 USD/month or 49 USD/yearSelf-hostedVerified Sep 16, 2026
4. Fluent
Fluent.ai's speech-to-intent engine maps spoken commands directly to device actions without transcribing to text first, which means no cloud round-trip, no NLP pipeline on a remote server, and no dependency on an internet connection. The technology runs embedded on low-power hardware and handles accent and language variation at the acoustic layer — not by training separate models per locale. Where it fits is narrow and deliberate: OEM device makers who need a voice interface that works in a noisy warehouse, a multilingual household, or a hearable that can't offload compute. Where it breaks is equally clear: if your use case needs open-ended conversation, dynamic vocabulary, or generative responses, this engine doesn't do that — it recognizes intent from a defined command set, not freeform speech.
PaidAPISelf-hostedVerified Jul 20, 2026
5. Mispher
Mispher runs speech-to-text and a lightweight local agent entirely on-device, targeting Apple Silicon Macs running macOS 26 and above. You dictate into any focused app field, issue spoken rewrite or translation instructions, or let the agent pull context from your screen, files, and notes — no packet ever leaves the machine. The MIT license means you can inspect, fork, and self-host without restriction. The ceiling arrives quickly: no API surface means integration into external pipelines requires custom code, and the agent's scope is bounded by what a local tool loop on a single Mac can reach.
FreeOpen SourceSelf-hostedVerified Jul 13, 2026
6. vaak — Speak. It types.
Vaak is an open-source desktop app that binds dictation to a hotkey and drops cleaned text into whatever app already has focus — your editor, CRM, terminal, or inbox, no workflow change required. You bring your own speech provider key, stored locally, so audio routes directly to the provider you trust, not through Vaak's servers. Filler words get stripped and punctuation gets added before the text lands. The ceiling appears fast: there is no API, no mobile support, and Windows is the primary download — macOS builds sit in GitHub Releases as early-preview artifacts. Teams who need server-side deployment, web app integration, or dictation across more than one OS will hit that wall quickly.
PaidOpen SourceSelf-hostedVerified Aug 14, 2026
7. Doculigent
Doculigent records screen and camera, streams a live diarized transcript for Teams, Zoom, and Google Meet by listening locally rather than injecting a bot, and hands you an AI summary before the tab is closed. Whisper runs on-device, which means your recordings and chat queries never leave the machine unless you choose cloud sync — and even then you can point it at your own S3-compatible bucket. The AI PM feature scans uploaded recordings across a team and rolls them into a single project status: blocked, at risk, on track. Where it strains: teams that need a hosted, zero-install experience for non-technical members will hit friction fast, and there is no public API, so anything downstream requires manual export rather than pipeline integration.
PaidSelf-hostedVerified Aug 16, 2026
8. Kami Subs
The pipeline is fixed and local: the browser extension captures tab audio, faster-whisper transcribes it, a translation layer converts it, and the result overlays directly on the video — no API keys, no per-minute billing, no audio leaving the device. It works on YouTube, Twitch, Vimeo, podcasts, and lecture streams, with one hard constraint: DRM-protected content is off-limits. The self-hosted backend means setup requires a working Python environment and a GPU capable of running faster-whisper at acceptable latency — that's a real installation step, not a one-click install. Community activity on the repository is minimal at the time of listing, so expect to self-diagnose when something breaks.
FreeOpen SourceSelf-hostedVerified Jun 1, 2026
9. KeyVox
KeyVox installs as a keyboard on iOS and a menu-bar app on macOS, transcribes via downloaded AI models (~190 MB for Whisper, ~480 MB for Parakeet), and writes text into any app the moment you release the trigger key. The Vibes feature applies reversible writing styles after dictation — so you can rephrase a casual voice note into a formal email without re-dictating. The Speak module reads copied text aloud using on-device voices. Where it breaks: KeyVox is a dictation keyboard, not a manual typing keyboard, so users who want to mix voice and touch typing in the same session have to switch back to their system keyboard constantly. Intel Mac support exists, but Apple Silicon is the architecture this runs well on.
PaidSelf-hostedVerified Jul 1, 2026
10. Voicetypr 2.0
Install it, pick a local Whisper or Parakeet model, bind a hotkey, and from that point forward a held key drops transcribed text into whatever app has focus — Gmail, Slack, Cursor, Notion, anything. No per-app configuration. The vendor states roughly 3× the words-per-minute of typing, and community feedback consistently flags offline speed as the standout surprise. Where it strains: the accuracy ceiling on local models is lower than cloud services, so dense technical jargon or heavy accents push users toward the optional cloud engines (Soniox, OpenAI, Groq, Deepgram). AI cleanup of rough dictation requires bringing your own API key — it is a paid-only feature that touches text only, never audio.
PaidFree Trial · 3 days$69 onceAPISelf-hostedVerified Jun 26, 2026
11. Whisper
Whisper solves the transcription bottleneck: turning audio from meetings, interviews, and podcasts into searchable text. It's trained on 680,000 hours of multilingual audio, so it handles accents and background noise better than most competitors. OpenAI charges $0.006 per minute of audio via API, with a free tier capped at modest monthly usage. The catch is real: heavy users quickly hit rate limits, and the free tier vanishes once you scale beyond hobbyist volume. You're paying per minute consumed, not per month.
FreeOpen SourceFree (open-source model)APISelf-hostedVerified Oct 1, 2023
12. Whissle Gateway
Whissle's Stream2Action architecture feeds audio, text, or video through a single-pass discriminative model — META-1 — and returns structured JSON carrying transcription, speaker diarization, emotion, intent, age, gender, and entities simultaneously. The full stack (ASR, LLM, TTS, diarization) runs self-hosted on a single GPU via Docker, which is the core production story here. The cloud API is documented as temporarily down while on-prem infrastructure is reinforced, so teams who need cloud failover have no fallback path right now. Video input is on a stated roadmap; text streaming arrives next. For contact center or privacy-sensitive workloads where you control the hardware, the on-prem path is active — for anything cloud-dependent, you are waiting.
PaidOpen SourceAPISelf-hostedVerified Jun 18, 2026
Listings on this page are sourced and verified by the AIDiveForge data pipeline. AIDiveForge is editorially independent — inclusion and rank are not for sale. Labeled ads are separate.