Streaming speech-to-text, three ways
Oruk has three WebSocket transcription paths. Realtime returns words as they are spoken in 32 locales, with an emotion score for each phrase. Hosted Orukeet returns final English text when an utterance ends, at $0.00045 a minute. Spectra-2 returns a transcript with emotion and style scores for clips in 25 languages.
Updated
Pick by what you need back
| Realtime | Orukeet (hosted) | Spectra-2 | |
|---|---|---|---|
| Model ID | oruk-realtime | oruk-orukeet | oruk-spectra-2 |
| Returns | Words as spoken, then a final transcript and an emotion score for each phrase | Final text when you commit the utterance | Final transcript with 15 emotion and 16 style scores |
| Languages | 32 locales, detected automatically | English | 25 languages, detected automatically |
| Length | Up to 10 minutes a session | Up to 60 seconds an utterance | Up to 60 seconds a clip |
| Speaker labels | Optional, for each phrase | Optional, priced separately | No |
| Price | $0.0080 a minute reference rate; uses plan minutes | $0.00045 a minute | Uses plan minutes |
| Status | Preview | Preview | Stable |
Plan minutes are the speech understanding allowance in each subscription. Hosted Orukeet has its own, larger allowance.
Open a Realtime session
import WebSocket from "ws"
const socket = new WebSocket(
"wss://speech-api.oruk.ai/v1/realtime?model=oruk-realtime",
"oruk-realtime",
{ headers: { Authorization: `Bearer ${process.env.ORUK_API_KEY}` } },
)
socket.on("open", () => {
socket.send(JSON.stringify({
type: "session.update",
session: { model: "oruk-realtime", language: "auto", sample_rate: 16000, phrase_emotions: true },
}))
})
// Mono signed 16-bit PCM, sent as binary frames while the person speaks.
export const send = (chunk) => socket.send(chunk, { binary: true })
export const finish = () => socket.send(JSON.stringify({ type: "input_audio_buffer.commit" }))
socket.on("message", (data) => {
const event = JSON.parse(data.toString())
if (event.type === "conversation.item.input_audio_transcription.delta") process.stdout.write(event.delta)
})What has been measured
Accuracy
Downloadable Orukeet: 9.85% pooled word error rate on 20,146 FLEURS recordings in 25 languages, down from 11.01% for the model it was tuned from.
Speed
JEV-Speech research runtime: warm 95th-percentile latency of 42.23 ms on one A100, down from 91.93 ms. A research measurement, not the hosted API.
Emotion
77.8 on speech-emotion-bench across 64,384 clips, first of 64 systems.
Questions
- Which path gives partial words while someone is still speaking?
- Only Realtime. Hosted Orukeet and Spectra-2 accept audio as it arrives and return one final result after you commit or finish.
- What audio does Realtime accept?
- Mono signed 16-bit PCM at 16 kHz, sent as binary WebSocket frames.
- How long can a stream run?
- A Realtime session lasts up to ten minutes. A hosted Orukeet utterance and a Spectra-2 clip are each limited to 60 seconds.
- Can I connect from a browser?
- Hosted Orukeet, yes: your server requests a 60-second, single-use token and the browser opens the WebSocket with it. For Realtime, connect from your server so the API key stays there.
- Is there a latency guarantee?
- No. Latency depends on your network, audio capture, language and phrase length. Measure end to end on your own traffic before you set a target.
Try it on your own audio.
The demo needs no account. The trial needs a card.