gRPC APIs · on prem, private cloud, or embedded

Speech engines you run yourself.

Every Cobalt engine is a protobuf service. Generate a client in your language, point it at your own server, and stream — the audio and the transcripts never leave your infrastructure.

Start with Transcribe Read the protos ↗

Engines · primary RPC

audio → text Transcribe

Speech recognition, streaming or batch, with word timings and confusion networks.

rpc StreamingRecognize(stream StreamingRecognizeRequest) returns (stream StreamingRecognizeResponse) audio → identity VoiceBio

Speaker enrollment, verification and identification from a voiceprint.

rpc StreamingIdentify(stream StreamingIdentifyRequest) returns (StreamingIdentifyResponse) text → audio VoiceGen

Streaming text to speech, synthesized as the text arrives.

rpc StreamingSynthesize(StreamingSynthesizeRequest) returns (stream StreamingSynthesizeResponse)

Integration path

Three steps to a working client

The order matters: you generate stubs from the proto, open a connection to your server, then stream. Transcribe is used here as the example; VoiceBio and VoiceGen follow the same shape.

step 1 Generate the client
Run buf against the proto definition to get native bindings for your language.
step 2 Connect to the server
Set the address, pick TLS or insecure, and confirm the server with Version and ListModels.
step 3 Stream audio
Send a config message, then audio frames, and read results as they come back.

Also in these docs

Beyond the engine APIs

Recognition context Bias Transcribe toward names, places and domain vocabulary.
Hybrid vs end-to-end Which model architecture fits your latency and accuracy budget.
Transcribe FAQ Accuracy, audio formats, languages and deployment questions.