Our Technology
Luqira trains its own avatar and voice models and runs them on its own stack. Published corpora: OpenSLR SLR94, SLR37, SLR41-44 and Google FLEURS.
We didn't wire up someone else's AI. We trained the model.
Most AI livestream tools are a thin wrapper around an API they rent. Luqira's avatar and voice models are our own, trained in house and served on our own inference stack. That is why we can answer in under a second, clone a voice from a short sample, and run hundreds of live streams at once from the same model.
Reaction to live comments, Native localized voices, Languages in the voice catalogue, Concurrent streams, one model, 100s
More than 85 languages, one training run.
Anyone can say they trained their own model. Here is what actually went into ours, named with corpus ids, so you can check it at the source. More than 85 languages, trained together in a single run, and 85+ of them reach you as output languages in the live voice catalogue.
Everything trains together in a single run, never one after another. Sequential passes overwrite each other until the model speaks only whatever it saw last. One mix, one run, one model. Every corpus id above is OpenSLR's own and was checked against OpenSLR's index rather than quoted from memory, so none of this has to be taken on trust. The full voice catalogue serves more languages still.
Every corpus in the mix is listed in full on the model card, with its id, licence, sample rate and hours.
Corpora
Mls
The speaker engine of the run. It adds no new language, since the model already spoke all eight, but thousands of distinct readers is what teaches a voice to stay itself instead of drifting toward an average.
Multilingual LibriSpeech (SLR94)
English, German, Dutch, French, Italian, Spanish, Portuguese, Polish
The cleanest audio in the mix, recorded at 48 kHz in studio conditions rather than scraped from the web. Small, and every hour of it teaches a language outright rather than polishing one already known.
Google studio TTS (SLR37, 41-44)
Bengali, Javanese, Khmer, Nepali, Sundanese
Google Indic
The largest genuine language-teaching block in the mix: six Indian languages plus Burmese, recorded to the same studio standard. Seven ids, not the unbroken SLR63-80 range it is often written as, and it carries no Hindi and no Punjabi — both of those reach the model through FLEURS instead.
Google crowdsourced speech (SLR63-66, 78-80)
Malayalam, Marathi, Tamil, Telugu, Gujarati, Kannada, Burmese
Bible
The highest value per hour anywhere in the mix. Twi was almost unknown to the model beforehand, and this brings audiobook-grade fidelity across West and Central African languages that most engines never attempt. Five of its six aligned languages are used, all but Ewe.
BibleTTS (SLR129)
Twi (Asante and Akuapem), Hausa, Lingala, Yoruba
Aishell
A voice pass on Mandarin, not a language lesson. The model was already fluent, so 400 speakers of clean indoor recording changes how it sounds rather than what it knows.
AISHELL-1 (SLR33)
Mandarin Chinese
Yoruba
The smallest corpus here, and a second pass on a language BibleTTS already teaches. Forty studio-clean hours of Yoruba from a different set of speakers is what stops one corpus deciding what a whole language sounds like.
Yoruba multi-speaker TTS (SLR86)
Yoruba
Fleurs
The breadth layer. Roughly ten hours each of parallel read speech, and the reason the run reaches past eighty-five languages instead of the twenty-five the OpenSLR corpora cover between them. Hindi and Punjabi arrive here rather than from the studio blocks.
Google FLEURS
102 languages, spanning and extending all of the above
Trained by us on OmniVoice. The weights are ours.
Luqira Voices is trained in-house on OmniVoice, an audio-codebook speech architecture built on the Qwen3-0.6B language model. Both are Apache-2.0, and we say so rather than implying we invented the architecture. What we trained is the model itself: the parameters start from random initialisation on that architecture, so no third-party model weights are inherited. We chose the corpora, built the mix, ran the training, and own the weights we trained, with every right to use them commercially and no licence to renegotiate with anyone, while Qwen and k2-fsa keep copyright in their own work. We rent no model from any provider. They run on our own GPUs and are never handed to a third-party API.
Our weights
Published model card, checkable off our own domain: huggingface.co/GuidenAI/genpio_voice
public corpora, hours pooled, languages, training run
Everything, Teaches new languages, Adds speakers, Voice quality pass, Adds breadth
capped
Bar length is logarithmic. Figures are raw hours available, before per-language caps.
Corpus size is not training share. Multilingual LibriSpeech brings roughly 1,250 times more audio than the smallest corpus in the mix, and per-language caps flatten that on purpose, so the languages actually being taught are never drowned out by the ones already fluent.
How our models are built and licensed
Yes, and we are specific about what that means. Luqira Voices is trained in-house from random initialisation on the Apache-2.0 OmniVoice architecture, itself built on Qwen3-0.6B. It is not a fine-tune of anyone's checkpoint: no third-party model weights are inherited, and the published model card is tagged trained-from-scratch. We chose the training corpora (OpenSLR SLR94, SLR37, SLR41-44, SLR63-66, SLR78-80, SLR86, SLR129, SLR33 and Google FLEURS, more than 60 languages in one run), built the mix and ran the training. Stated precisely, because both halves are checkable: we own the weights we trained and hold every right to use them commercially, under a permanent Apache-2.0 grant over the upstream architecture, with no field-of-use limit, no revenue cap and no licence to renegotiate with anyone, while Qwen and k2-fsa keep copyright in their own work. We rent no model from any provider and resell no third-party text-to-speech. Everything runs on our own inference stack with no third-party API in the serving path, which is why we can hold latency, cost and quality where a reseller cannot. The model card is public at huggingface.co/GuidenAI/genpio_voice, so none of this has to be taken on our word.
Rented AI has a ceiling. Ours doesn't.
You have seen the mix and the licence. Here is what owning the weights changes, line by line, against a tool that rents the same capability from someone else.
Wrapper tools
Luqira
Rows
Cost
Pay a vendor by the minute, and watch the bill grow with every stream
We own the inference, so adding streams doesn't multiply the cost
Latency
Latency is whatever the vendor hands you that day
We tune latency ourselves, down to the frame
Quality
A quality bug is a support ticket and a long wait
A quality bug is something we fix in the model
Scale
One session at a time, one rate limit away from dark
Add the hundredth concurrent stream without asking anyone: the GPUs are ours
Built by us, so it bends to you
Your face, rendered live.
Clone your likeness or start from a branded host. Lip-sync is generated frame by frame, not stitched from clips, so the mouth follows the words instead of chasing them.
Your voice, from a short sample.
A cloned voice that keeps your cadence and your warmth. Across 600+ localized voices it reads a price, a name and a punchline the way a native speaker would, not the way a translation sounds read aloud.
Knows your catalog cold.
Import products, pricing, FAQs, and selling points. The host answers accurately and stays on message, even when a viewer pushes.
Every platform, at the same time.
TikTok, YouTube, Shopee, Instagram, and more, live in parallel from one workspace.
A comment becomes a sale in under a second
Four stages, one loop, running for as long as the stream stays live.
under 1s, end to end
The answer is what provokes the next comment, so the loop starts again.
Listen
Comments, questions, and reactions arrive from every connected platform at once.
Understand
Intent, sentiment, and buying signals are read against your catalog, your pricing, and your brand voice.
Speak
Our voice model answers in your cloned voice, in the language the viewer is actually typing in.
Show
Our avatar model renders the reply lip-synced frame by frame and pushes it live.
Set your stream up by asking an AI assistant.
Luqira speaks MCP, the protocol AI clients use to reach outside tools. Connect it once and your assistant can open your account, size your plan, write your scripts and fill your product library, so your first stream starts as a conversation instead of an afternoon of setup.
- Point your client at Luqira — One endpoint, no API key to create. Paste it into ChatGPT, Claude, Codex or whichever client you already use.
- Approve it in your browser — The assistant hands you a link and a short code. You sign in on our own page and approve it there. It never sees your password.
- Ask for what you need — An account, the right plan, tonight’s script, products read straight from a shop link. It does the setup while you talk.
Chat
Any MCP client
Set Luqira up for tonight and add these three products.
Open this link and approve code 4F2C-91KD, then I will take care of the rest.
Account connected
Plan sized, checkout link ready
Opening script written and saved
3 products imported from your links
It never sees your password — Signing in happens on our page, in your own browser.
It never holds your credentials — Connecting returns a handle that means nothing outside Luqira.
It never charges your card — It can hand you a checkout link. You complete the payment yourself, on the secure checkout page.
Infrastructure you can count on
Enterprise-grade security
TLS 1.3 in transit, AES-256 at rest, and voice cloning that will not run without the speaker’s consent. Every control is listed in full on our security page.
Read the security page
99.9% uptime SLA
A committed number on multi-region infrastructure with automatic failover, not an aspiration. Global edge delivery keeps latency low when revenue depends on the stream.
Open API and integrations
Wire Luqira into your CRM, OMS, analytics, and storefront. The API reference is public and needs no key to read.
Read the API reference
Dedicated support
Priority onboarding, training, and an account team that owns the outcome with you.