Meetings, calls, dictation, depositions: the recordings that most need transcription are exactly the ones nobody should upload. Transcribe them where they were recorded.
Every organization now records itself: standups, client calls, interviews, clinical dictation. What happens next is either a privacy decision or a privacy accident.
The same engine that transcribes also detects speech, identifies the language, translates, and summarizes, so the pipeline never leaves the process.
03
Segments stream as they decode, timestamped, ready for captions or notes.
The scenario decides the form: a transcription service the whole team calls, or transcription living inside your own application.
Serve it
One server transcribes for every tool that can POST a file: meeting archives, call recordings, media assets. Authenticated, audited, governed by policies you set, and priced by nothing per minute.
Embed it
Ship dictation, captions, or note-taking as a feature of your own .NET application, including offline and edge installs where the network is not part of the plan.
The properties a privacy officer asks about first, stated plainly.
Private
Recording, transcript, and every derived artifact stay on hardware you control; air-gapped operation works once models are present.
Live
Segments arrive as they decode, so captions, dictation, and live translation feel immediate instead of after-the-fact.
Unmetered
Backfilling years of recordings costs hardware time, not per-minute fees; the economics favor exactly the heaviest users.
The served form is one HTTP call; the embedded form is one class with a streaming event.
POST a recording to LM-Kit One from any language or tool.
curl http://your-server:5189/lmkit/v1/audio-transcription \
-H "Authorization: Bearer your-server-token" \
-F "file=@weekly-sync.mp4"The embedded engine streams timestamped segments as they decode.
using LMKit.Model;
using LMKit.Speech;
var model = LM.LoadFromModelID("whisper-large-v3-turbo");
SpeechToText engine = new(model);
// Segments stream in as they decode: captions, notes, or storage.
engine.OnNewSegment += (sender, e) => transcript.Append(e.Segment.ToString());
engine.Transcribe(audio);The pipeline composes documented capabilities you can inspect on their own.
Transcribe
Streaming speech-to-text on the Whisper family, timestamped segment by segment.
How transcription worksIsolate
Speech separated from silence and noise before a token is spent on it.
How VAD worksTranslate
Live translation while transcribing, for multilingual meetings and calls.
How speech translation worksCondense
Transcripts become minutes and action lists with the same local engine.
How summarization worksRecall
Indexed transcripts answer "what did we decide" with the meeting and minute cited.
How RAG chat worksRelated
The use case transcripts feed: one searchable corpus for everything the team knows.
The knowledge use caseYes. The Whisper-family models run on your own CPU or GPU through LM-Kit, so the recording, the transcript, and every summary or translation derived from it stay on hardware you control. Air-gapped operation works once the models are present.
Yes. The spoken language is detected automatically, transcription covers the Whisper family's multilingual range, and speech translation can render segments into another language as they decode.
Deploy LM-Kit One and POST recordings to its transcription endpoint from any tool or language; the server authenticates callers, applies your policies, and logs usage. Embedding with LM-Kit.NET is for putting transcription inside your own application.
There is no per-minute meter. Throughput is set by your hardware and the model you choose, so backfilling an archive of recordings is a scheduling decision rather than a budget event.
Private AI, spoken