Logo
Miralas Voice Intelligence

Voice AI, in progress.

Research updates, model comparisons, language expansion and real voice samples from the Miralas team.

Native-language researchMultilingual TTSVoice cloningRealtime evaluation
Miralas baseline
500M
Chatterbox Multilingual V3
Baseline coverage
23+
Chatterbox multilingual languages
Audio
Realtime
Voice AI evaluation stack
Uzbek
Native
Miralas training direction

Latest Updates

Research•August 31, 2026•5 min read

Why Uzbek Is a First-Class Training Track for Miralas

Instead of treating Uzbek as a translation afterthought, Miralas is building language-specific data, phoneme coverage and evaluation around native speech.

Read update
Multilingual•August 28, 2026•4 min read

Adding More Global Languages to the Evaluation Lab

English, Spanish, Chinese, Hindi, Arabic, Japanese, Korean, French, German, Portuguese, Turkish and Russian are now part of the broader comparison set.

Read update
Models•August 25, 2026•6 min read

Miralas vs. the New Generation of Voice AI

We are comparing Miralas with GPT-Realtime, Gemini Live, Grok Voice and the Chatterbox baseline using transparent capability categories rather than invented leaderboard numbers.

Read update
Engineering•August 22, 2026•7 min read

Inside the Miralas Training Pipeline

From clean speech data and speaker embeddings to evaluation and inference, this is the direction behind our next voice models.

Read update

Voice Samples

Side-by-side reference audio for the models we evaluate. Replace the local files with your own generated or licensed samples before shipping.

Miralas

Miralas / Chatterbox Multilingual V3

Miralas

Open multilingual baseline with Miralas-specific training and evaluation.

Miralas

Miralas / Chatterbox Multilingual V3

Audio sample
0:00 / 0:12
Voice cloningMultilingual TTSCustom trainingUzbek research

GPT-Realtime

OpenAI

OpenAI

Realtime text/audio model for conversational voice applications.

GPT-Realtime

OpenAI

Audio sample
0:00 / 0:12
Realtime audioAudio input/outputWebRTCWebSocket / SIP

Gemini 3.1 Flash Live

Google

Google

Google's realtime audio model for natural dialogue through the Live API.

Gemini 3.1 Flash Live

Google

Audio sample
0:00 / 0:12
Realtime dialogueAudio input/outputLive APIMultimodal

Grok Voice

xAI

xAI

xAI voice capability included as an external comparison reference.

Grok Voice

xAI

Audio sample
0:00 / 0:12
VoiceRealtimeConversational AIExternal reference

Model Capabilities

A transparent capability map — not a fabricated cross-vendor benchmark. Latency, MOS, WER and reasoning scores will be shown only after Miralas runs the same evaluation methodology across every system.

Realtime audio✓✓✓✓
Audio input / output✓✓✓✓
Voice cloning✓———
Open-source baseline✓———
Custom language training✓———
Uzbek training track✓———
Miralas baseline
500M
Chatterbox Multilingual V3
Chatterbox coverage
23+
Languages out of the box
GPT-Realtime
Audio I/O
Realtime voice model
Gemini Live
A2A Audio
Gemini 3.1 Flash Live
Benchmark policy: vendor capabilities are presented as documented capabilities. Internal Miralas metrics will be labeled separately with the evaluation date, hardware, dataset and methodology.

Languages

Popular global languages are part of our evaluation set. Uzbek has a separate Miralas-native training direction.

🇬🇧Baseline

English

en

🇪🇸Baseline

Spanish

es

🇨🇳Baseline

Chinese

zh

🇮🇳Baseline

Hindi

hi

🇸🇦Baseline

Arabic

ar

🇯🇵Baseline

Japanese

ja

🇰🇷Baseline

Korean

ko

🇫🇷Baseline

French

fr

🇩🇪Baseline

German

de

🇵🇹Baseline

Portuguese

pt

🇹🇷Baseline

Turkish

tr

🇷🇺Baseline

Russian

ru

🇺🇿Miralas native training

Uzbek

uz

Uzbek — Miralas Native Training Track

Uzbek is being treated as a first-class training and evaluation target rather than simply relying on a generic multilingual checkpoint. The program focuses on native pronunciation, phoneme coverage, prosody, speaker consistency and real-world conversational speech.

Native dataSpeaker evaluationPhoneme coverageProsody