Speech infrastructure for companies and creators
Voice in.Voice out.Everything in between.
Clean, transcribe and generate speech in Studio without code. Run voice agents on your phone lines and in your app. Transcribe and denoise audio from your own code with the API.
01Clarity
Caller audio
rawclean
02Scribe
Transcript
Hi, I need to move my appointment to Thursday.
03Agent
Decision
intentreschedule_appointmentknowledgeClinic hours
Sure. I have Thursday at 10:30 or 2:15. Which works better?
04Narrate
Reply audio
voice · Mara
02Models
Four models.One turn of a voice agent.
The same Northside Dental sample call, from the caller’s audio to the spoken reply, one model at a time.
- InCaller audio, phone lineSample call · Northside Dental
- 01
Clarity
Speech enhancement
Latency about 10–15 ms
Clean, natural voice from noisy recordings.
APIclient.denoise.process()Clarity · voice keptPhone line · raw - 02
Scribe
Speech transcription
Latency about 120 ms
Speech to text in 18 languages, or let it detect the language.
APIclient.asr.transcribe()- 00:00.34Caller
Hi, I need to move my appointment to Thursday.
Caller word timings00:00.34–00:02.86
- 00:00.34Caller
- 03
Agent
Conversation model
Understands the caller and decides the reply in 18 languages.
- heard
- “Hi, I need to move my appointment to Thursday.”
- intent
- reschedule_appointment
- details
- day = Thursday
- knowledge
- Clinic hours→Thu 10:30 · 2:15
- decides
- Offer two Thursday slots→reply
End of turn 00:02.86Decided 00:03.10
- 04
Narrate
Speech generation
Latency about 350 ms
Turns plain text into spoken audio in 18 languages.
Voice · Mara
Sure. I have Thursday at 10:30 or 2:15. Which works better?
Reply audio00:03.30–00:06.62
- OutReply audio, back to the callerFull turn under 1 s
03API
Transcribe and denoise from your own code.
Speech to text in 18 languages, background noise removal and storage, through one TypeScript SDK or the REST API.
import { readFileSync } from 'fs';
import { Decrackle } from '@decrackle/sdk';
const client = new Decrackle({
apiKey: process.env.DECRACKLE_API_KEY,
});
const audio = readFileSync('./call-0142.wav');
// Transcribe one call; the language is detected
const result = await client.asr.transcribe(audio);import { readFileSync } from 'fs';
import { Decrackle } from '@decrackle/sdk';
const client = new Decrackle({
apiKey: process.env.DECRACKLE_API_KEY,
});
const audio = readFileSync('./call-0142.wav');
// Remove background noise, keep it as FLAC
const clean = await client.denoise.process(audio, {
outputFormat: 'flac',
});{
"jobId": "job_0142",
"status": "completed",
"detectedLanguage": "en",
"audioMetadata": { "durationSeconds": 2.86 },
"transcriptionText":
"Hi, I need to move my appointment to Thursday."
}{
"jobId": "job_0143",
"status": "completed",
"outputUrl": "https://…/call-0142.flac",
"outputFormat": "flac",
"outputEphemeral": false,
"audioMetadata": { "durationSeconds": 2.86 }
}04Why Decrackle
One voice stackinstead of four vendors.
One contract, one bill, one security review.
< 1 s
Response time, end of caller speech to first reply audio
18 languages
Scribe, Narrate and Agents in English, German, French, Spanish, Italian, Portuguese, Swedish, Danish, Hindi, Gujarati, Marathi, Tamil, Telugu, Bengali, Japanese, Thai, Nepali and Hebrew
Three ways to use the models
Decrackle Studio
No-code apps for creators, media teams and call centers.
Content Creator, Conversation Intelligence
Decrackle Agents
Voice agents on phone calls and in your app, grounded in Knowledge Bases, guided by Workflows.
Phone lines, in-app voice
Decrackle API
Speech to text, noise removal and storage from your own code, through a TypeScript SDK or REST API.
asr, denoise, storage
05FAQ
Questions,answered.
How fast is a Decrackle voice agent?
A full voice-agent turn takes under 1 second, from the end of the caller’s speech to the first audio of the reply. Each model is fast on its own too. Clarity, which cleans the audio, has a latency of about 10 to 15 ms. Scribe, which transcribes it, takes about 120 ms, and Narrate, which speaks the reply, about 350 ms.
Which languages can Decrackle transcribe?
Scribe, our speech to text model, supports 18 languages. They are English, German, French, Spanish, Italian, Portuguese, Swedish, Danish, Hindi, Gujarati, Marathi, Tamil, Telugu, Bengali, Japanese, Thai, Nepali and Hebrew. Narrate and Agents support the same 18. Pass the language as an ISO code, such as "en", or as its full name, such as "English". If you leave it out, the API detects the language by itself and returns it with the transcript.
What can I build with the Decrackle API?
The API covers speech to text, audio denoising and storage. You call it through the official TypeScript SDK, @decrackle/sdk, which runs on Node.js 18 or later, or through the REST API. Every method is asynchronous and returns a typed result, errors arrive as typed error classes, and past jobs can be listed, fetched and deleted.
Which audio formats does Decrackle accept?
For speech to text, the API accepts MP3, WAV, FLAC, OGG, M4A and AAC files. Denoising also takes AIFF, OPUS and WEBM. You choose the format of the cleaned file when you upload it, WAV, MP3, FLAC or Opus, depending on your plan. The format cannot be changed after the job is created, so pick it before you start.