Speech AI Platform

Speech models built for how Southeast Asia actually speaks.

Aisyah understands speech, Nada generates it. Both trained in-house on Bahasa Melayu, Manglish, regional accents and the everyday switching between Malay and English that generic models handle badly.

Access opens with free minutes on the flash models. No card required.

Transcribe a file
curl -X POST https://api.revolab.ai/v1/stt \
  -H "Authorization: Bearer $REVOLAB_API_KEY" \
  -F "model=aisyah-1.0-pro" \
  -F "file=@call-recording.wav"
Two endpoints: /v1/stt to transcribe, /v1/tts to speak.
5.79%

Word error rate, Aisyah 1.0 Pro. Most accurate of the fourteen models tested.

0.62%

Invented words. The lowest of all fourteen, including models it beats on accuracy.

300 ms

To first audio from Nada 1.0 Pro. A blink takes about 400.

1,899

Real Malaysian utterances. Every model transcribed the same set.

The Models

Two families. One understands speech, one produces it.

Speech to text
Aisyah 1.0 Pro
aisyah-1.0-pro

The accuracy tier, for messy, real-world, code-switched Malaysian calls. Most accurate of the fourteen models tested, and the least likely to invent words that were never said.

Word error rate
5.79%
Invented words
0.62%
Rank
1 of 14
Speech to text
Aisyah 1.0 Flash
aisyah-1.0-flash

The fast tier, for live transcription where latency matters more than the last decimal point. Still ahead of Gemini 2.5 Flash and Whisper large-v3.

Word error rate
10.03%
Invented words
0.72%
Rank
6 of 14
Text to speech
Nada 1.0 Pro
nada-1.0-pro

Malay-first voices that speak a sentence switching between Malay and English without stumbling. Audio starts before the sentence is finished, so a live agent answers without dead air.

Time to first audio
300 ms
Streaming
Yes
Voice cloning
Yes
The Benchmark

The number that breaks automation is not accuracy.

Word error rate is how much of what was said came back wrong. Invented words are text the model produced that nobody spoke, and that is the number that does the damage: it flows straight into call summaries, routing decisions and compliance logs. Aisyah 1.0 Pro has the lowest invented-word rate of all fourteen models, including ones it beats on accuracy. Lower is better in both columns.

Word error rate and invented-word rate across fourteen models
#ModelWord error rateInvented words
1Aisyah 1.0 Pro5.79%0.62%
2Scribe v26.67%0.70%
3Gemini 2.5 Pro7.00%1.00%
4Qwen-Audio 3.0 ASR Flash9.17%1.11%
5Ilmu ASR v4.29.48%0.87%
6Aisyah 1.0 Flash10.03%0.72%
7Gemini 2.5 Flash10.76%2.48%
8Whisper large-v317.30%2.73%
9Gemini 3.6 Flash24.55%14.35%
10Nova 331.99%0.94%

Ten of fourteen models shown. Benchmarked on 1,899 real Malaysian utterances, 820 published and 1,079 held back. Aisyah 1.0 Pro scores 4.89% on the published half and 6.58% on the held-back half, so its accuracy holds up on audio it has never seen.

What It Does

Built around Malaysian speech from the start.

Speaker diarization

Set diarize and every segment comes back tagged with who spoke it. An agent and a customer, or three people on a conference bridge, arrive as separate speakers instead of one run-on block of text.

Two languages in one sentence

Code-switching between Malay and English is a first-class setting, not a workaround. Turn on language tags and each segment is labelled with the language it was actually spoken in.

Denoising on the way in

Real call audio is compressed, noisy and narrowband. Denoising is a request flag on the same endpoint, ahead of a voice-activity pipeline that splits long recordings into segments.

Speech that streams as it is generated

Set stream and Nada returns raw PCM as it synthesizes, so a live agent starts speaking before the sentence is finished. Billing meters only the audio actually delivered.

Voice cloning

Clone a voice from a short reference sample through a guided wizard, then use it like any premade voice. A pronunciation dictionary lets you fix how specific words are said.

The audio you already have

WAV, MP3, M4A, FLAC, OGG, Opus and WebM, up to 50 MB or 30 minutes a request. Give it a language hint or let it detect one.

Migration

Keep the SDK you already use.

The interface follows the OpenAI convention, so existing SDKs work by pointing at a different address. The gateway also speaks the ElevenLabs and Cartesia APIs, and LiveKit agents connect through the stock Cartesia plugin.

OpenAI
ElevenLabs
Cartesia
LiveKit
Enterprise

Or run it inside your own walls.

The models already run on our own GPU servers rather than someone else's API, so the same stack can be deployed in your environment when audio is not allowed to leave it. Banks, telcos and government work usually start here.

Published pricing is being finalised. Tell us the volume and how you need it deployed, and we will quote it.

Talk to the team
In the platform today
  • Two-factor authentication, required on every account
  • API keys scoped to speech to text, text to speech, or both
  • Optional per-key expiry, and revocation that takes effect immediately
  • Every record in the platform belongs to an organisation
  • Your audio is never used for training without an explicit opt-in
With an enterprise agreement
  • Deployment inside your own environment
  • Single sign-on
  • Role-based access across your organisation
  • Pricing scoped to your volume

Scoped with our team rather than switched on at signup.

Run your own audio through it.

The waitlist is open now. Access opens with free minutes on the flash models, no card required. If your recordings are Malaysian calls, that is the test worth running.