Speech-to-Text - Lokutor

Standalone transcription — no LLM turn, no TTS, no voice-agent session. Use STTClient for a single pre-recorded clip, or SpeechToTextClient for continuous/live transcription from a microphone.

Batch transcription

STTClient sends one complete audio clip to POST /stt/transcribe and returns the transcript.

Node.js
Browser
Python

import { STTClient, Language } from '@lokutor/sdk';
import fs from 'fs';

const client = new STTClient({ apiKey: process.env.LOKUTOR_API_KEY! });

const result = await client.transcribe({
  audio: fs.readFileSync('recording.wav'),
  language: Language.ENGLISH,
});

console.log(result.text);

import { STTClient } from '@lokutor/sdk';

const client = new STTClient({ apiKey: import.meta.env.VITE_LOKUTOR_API_KEY });

// From a <input type="file"> or a MediaRecorder Blob
async function transcribeFile(file: Blob) {
  const result = await client.transcribe({ audio: file });
  console.log(result.text);
}

from lokutor import STTClient, Language
import os

client = STTClient(api_key=os.environ['LOKUTOR_API_KEY'])

with open('recording.wav', 'rb') as f:
    result = client.transcribe(f.read(), language=Language.ENGLISH)

print(result['text'])

Options

Option	Type	Default	Description
`audio`	`Blob \| ArrayBuffer \| Uint8Array` (JS) / `bytes` (Python)	required	A complete WAV file, or raw PCM16 with `format: "pcm16"`
`format`	`"wav" \| "pcm16"`	`"wav"`	Only needed for raw PCM16 input
`sampleRate` / `sample_rate`	`number`	—	Required when `format` is `"pcm16"`; ignored for WAV (read from the file header)
`language`	`Language`	server default	ISO language code

Result

Field	Type	Description
`text`	`string`	Full transcript
`latencyMs` / `latency_ms`	`number`	Server-side processing time
`engine`	`string`	STT engine used
`sampleRate` / `sample_rate`	`number`	Sample rate of the processed audio
`language`	`string`	Detected/used language
`durationSeconds` / `duration_seconds`	`number`	Audio duration
`segments`	`array`	Per-segment text with timing, when the engine provides them

Live transcription

SpeechToTextClient streams microphone audio to WS /ws/stt. The server’s VAD detects utterance boundaries automatically — you get partial transcripts while the user is speaking and a final transcript once they stop.

Browser
Node.js
Python

import { SpeechToTextClient, Language } from '@lokutor/sdk';

const client = new SpeechToTextClient({
  apiKey: import.meta.env.VITE_LOKUTOR_API_KEY,
  language: Language.ENGLISH,
  onPartialTranscript: (text) => console.log('...', text),
  onFinalTranscript: (text) => console.log('Final:', text),
  onError: (err) => console.error(err),
});

await client.connect(); // prompts for mic permission

// later
client.disconnect();

import { SpeechToTextClient, NodeAudioManager, Language } from '@lokutor/sdk';

const client = new SpeechToTextClient({
  apiKey: process.env.LOKUTOR_API_KEY!,
  language: Language.ENGLISH,
  onFinalTranscript: (text) => console.log('Final:', text),
});

await client.connect(new NodeAudioManager());

from lokutor import SpeechToTextClient, Language
import os, time

client = SpeechToTextClient(
    api_key=os.environ['LOKUTOR_API_KEY'],
    language=Language.ENGLISH,
    on_final_transcript=lambda text: print('Final:', text),
)
client.start()

time.sleep(60)
client.stop()

Options

Option	Type	Default	Description
`apiKey` / `api_key`	`string`	required	—
`language`	`Language`	`ENGLISH`	ISO language code
`vad`	`"silero" \| "rms"`	`"silero"`	VAD engine: neural (default) or energy-threshold fallback
`onPartialTranscript` / `on_partial_transcript`	`(text) => void`	—	Fires repeatedly mid-utterance with the best guess so far
`onFinalTranscript` / `on_final_transcript`	`(text) => void`	—	Fires once the utterance ends
`onError` / `on_error`	`(error) => void`	—	—
`onStatusChange`	`(status) => void`	—	JS only: `"connecting" \| "connected" \| "disconnected"`

Call endUtterance() / end_utterance() to force-finalize whatever’s currently in progress (e.g. a “done speaking” button) instead of waiting for VAD silence.

Which one do I want?

A file the user already recorded or uploaded → STTClient.transcribe().
Live captions, dictation, or transcribing what’s coming from a microphone right now → SpeechToTextClient.
Building a full voice conversation (LLM responds, TTS speaks back) → VoiceAgentClient instead — it already includes STT as part of the turn.