Turn live conversations and recorded audio into timestamped transcripts without sending them to a cloud API. Run on your own CPU or GPU infrastructure, with optional translation and a typed API for your existing systems.
Talk directly with an engineer about your audio, hardware and security boundary.
From audio input to transcript output, processing stays inside your environment. Illustrative flow; capacity is evaluated for your hardware and workload.
Turn live audio and recorded files into timestamped transcripts inside your infrastructure.
Work inside restricted networks, with offline installation and no cloud API dependency.
Connect your existing systems through typed gRPC calls for batch and streaming audio.
Conventional cloud transcription is convenient until the recordings are confidential. Then every upload becomes a question for legal, security and finance. gensft Speech brings speech recognition to the place your recordings already live. Make conversations searchable and usable without handing the audio to an external transcription service.
Audio and transcripts remain inside your controlled environment. Nothing is uploaded, and nothing is retained unless you configure it to be.
Deploy in private networks, data centres, restricted environments or fully air-gapped infrastructure. Internet access is not a requirement.
Bounded workloads, admission control and hardware-specific capacity profiles give predictable behaviour under demand — no surprise bills, no silent degradation.
Strongly typed APIs for batch processing, bidirectional live streaming, health monitoring and capability discovery slot into the software you already run.
Receive recorded files or live audio streams over the API.
Decode, resample, validate and prepare the audio for inference.
Process speech using locally deployed inference workers on CPU or GPU.
Optionally translate completed transcript segments in place.
Return ordered, timestamped partial or final results through the stream.
A speech model is a component. A speech service has to accept a thousand malformed files, run for months, refuse work it cannot finish and shut down cleanly when told. gensft Speech is written in modern C++ with that second job in mind.
Deploy in private networks or fully air-gapped environments, with CPU and NVIDIA GPU options for your workload and offline installation packages.
For environments prioritising compatibility, simplicity and controlled infrastructure costs. Runs on the servers you already have.
For higher throughput, lower latency and larger concurrent workloads. Capacity profiles are published per supported GPU.
Built for environments where cloud processing is not an option. Every control below is a property of the software you run, verifiable on your own hosts.
Regulatory compliance depends on your deployment, configuration, policies and operating environment. gensft Speech provides controls that support a compliant deployment; it does not by itself constitute certification under any regime.
Board, deal and HR conversations become searchable minutes without a recording ever leaving the room's network.
Recorded statements are transcribed with timestamps that hold up to review, under chain-of-custody your team controls.
Clinical dictation and consultations are captured on-site, so patient audio stays with the institution that recorded it.
Hearings, consultations and casework recordings are processed inside the department's own accredited environment.
Every call is transcribed for quality and compliance review at a fixed infrastructure cost, not a per-minute one.
Interview corpora are transcribed and translated in bulk, without participant audio being shared with a third party.
Decades of broadcast audio become text you can search, on hardware sized to the backlog.
Teams working across languages get translated transcripts from a system that is never reachable from the internet.
Performance depends on the selected model, quantisation, hardware, audio quality, language and concurrency. Figures are published per hardware profile once measured on reference workloads.
gRPC with Protocol Buffers. Batch calls for files, a bidirectional stream for live audio, and health and capability endpoints so your orchestrator knows what a node can do before it sends work.
service Speech {
rpc Transcribe(TranscribeRequest) returns (Transcript);
rpc StreamTranscribe(stream AudioChunk)
returns (stream TranscriptEvent);
rpc GetCapabilities(Empty) returns (Capabilities);
rpc Check(HealthRequest) returns (HealthStatus);
}
message TranscriptEvent {
uint32 segment_id = 1;
bool is_final = 2;
int64 start_ms = 3;
int64 end_ms = 4;
string text = 5;
string translation = 6; // optional
}No. Ingest, inference, translation and delivery all run on the hosts you deploy. The service makes no outbound connections as part of processing.
Bring your deployment constraints. We’ll discuss how gensft Speech fits your infrastructure and what to evaluate before you commit.
Private network, on-premise or air-gapped requirements.
Languages, live streams, batch volumes and available hardware.
Where transcripts go and how the API fits your existing systems.
Tell us your name, organisation, languages, and expected workload. Tell us where you plan to deploy and about any hardware or integration requirements.
Your details are stored privately for our team to review. Please don’t include passwords or sensitive documents. Prefer email? hello@gensft.com