Ganito actually magsalita ang mga customers mo.
Bakit hindi ang voice agent mo?
Two billion people switch languages mid-sentence. Vaakya builds self-hosted small language models, fine-tuned on real code-mixed sales calls, so your voice agents do too. One GPU. Nothing leaves your network.
The same sentence, three ways. Only what the model hands the voice engine changes.
Frontier models struggle on code-mixed utterances: dropping words, hallucinating grammar, or translating prematurely. Vaakya processes both linguistic layers natively, yielding reliable word-for-word transcripts. Even at conversational speed, under background noise, on low-bitrate phone lines.
इसकी शुरुआती क़ीमत 1.38 करोड़ rupees से है। क्या मैं project की brochure भेज दूं?
How a TTS reads it: every Hindi word gets Hindi phonetics, every English word gets English phonetics. करोड़ is a Hindi crore; brochure is an English brochure.
Sounds like a bilingual person.
Two ways to get a model that speaks like your customers
Deploy our ready-to-run foundation models or commission a private domain model trained on your proprietary audio with legal indemnification.
Private models
Fine-tuned to your domain under NDA.
| First delivery | 14 days to first functional checkpoint |
| WER | Sub-4% in domain |
| Rights | Full copyright & weight ownership assigned to you. We don't retain a copy. |
| Hosting | Air-gapped on your private cloud, VPC, or on-prem. No external API calls. |
- Custom acoustic lexicon for your company's slang, acronyms, and product names
- Synthetic data generation for low-frequency edge cases
- Human-in-the-loop validation by native bilingual speakers
Open-core models
Per-second open-source ASR, run anywhere, no telemetry, Apache 2.0.
| Inference time | Sub-100ms streaming chunk latency. Zero cloud dependency. |
| Architecture | Conformer-CTC streaming encoder. Quantized INT8 weights available. 180MB footprint. |
| Languages | Spanish-English, Hindi-English, Tagalog-English, Arabic-French, and 8 more pairs. |
| License | Apache 2.0. Completely free for commercial and non-commercial use. |
- Plug-and-play Docker image with zero dependencies. Runs on commodity GPUs or CPUs.
- Pre-integrated with standard telephony protocols (SIP, RTP, WebRTC via FreeSWITCH and LiveKit).
- Permissive open-source, deploy without licensing fees.
Every private model is delivered with a zero-retention guarantee. Models run entirely within your perimeter. No audio, transcripts, or telemetry ever leave your infrastructure.
Frontier models are trained on one language at a time. Your customers aren't.
On code-mixed calls, general-purpose LLMs fail in three specific, measurable ways. Each one is the moment a customer hears a machine.
Wrong script
An English word comes out in the local script — प्रोजेक्ट instead of project — or, in Spanish–English, goes untagged so the voice engine guesses. Either way the phonetics are wrong.
Corrupted names
The same product name is rendered four different ways across four calls. On a sales call, the name is the thing you're selling.
Unnatural switching
The model switches languages where no speaker would. A Hindi speaker says actually in English and लेकिन in Hindi; a Spanglish speaker says the meeting but mañana. That rhythm is learned from real calls, not written in a prompt.
What every model comes with
Ready or custom, hosted or on-prem: a language model for your voice agents. Not a calling platform, not telephony, not speech-to-text. It plugs into whatever you already run.
A voice that sounds like your customers' neighbour
Two languages mixed the way people in that market actually speak, every word in the right script or tagged for the voice engine, product names never mangled. Your existing TTS sounds human because the text it receives is.
A model that runs inside your walls
Small enough for one GPU. Deployed in your VPC or on-prem, in any region. No audio or transcript leaves your network, so the pilot survives legal review in Mumbai, Miami or Manila.
Numbers on your own calls
Every version ships with a scorecard: script fidelity, entity preservation, switch-point accuracy, latency. Measured against what you use today.
An answer for the regulator
Pinned model versions and logged inputs and outputs. When someone asks why the agent said that, there is a record.
Deployment
The whole conversation loop stays inside your network boundary.
one GPU
Caller ↔ your telephony ↔ this loop. Nothing crosses the dashed line.
- Runs on a single A10/L4-class GPU with sub-300 ms turn latency.
- Built for data-residency regimes: GDPR, HIPAA and SOC 2 controls in the US and EU; DPDP Act and RBI localisation in India.
- No per-token cost. Your bill does not grow with call volume.
- Works with Vapi, Retell, Bolna, in-house stacks, or any pipeline that passes text.
We publish where we win and where we lose
The first open benchmark for code-mixed voice generation. Real held-out sales calls, scored on the properties that make or break a conversation.
| Metric | Vaakya 8B, fine-tuned | Frontier API, best prompt + glossary | What it measures |
|---|---|---|---|
| Script / tag fidelity | — % | — % | Each word in the right script, or correctly language-tagged, per token |
| Entity preservation | — % | — % | Product and place names rendered correctly |
| Switch-point accuracy | — % | — % | Language switches where a human would |
| Naturalness (MOS, 1–5) | — | — | Native speakers rating TTS output blind |
| P95 turn latency | — ms | — ms | Text in to text out, single GPU vs API round-trip |
Methodology is public now. First results on Hindi–English real-estate calls publish in the coming weeks; insurance and Spanish–English results follow with our design partners.
One method, every code-mixed market
Each new pair is a data-collection problem, not a research problem. India was first because that is where our production calls came from.
Hinglish
Live. Trained on production real-estate sales calls.
Spanglish
In progress. US insurance and healthcare intake; per-word language tags for the voice engine.
Taglish
Planned. Philippines BPO and insurance.
Arabic-French
Planned. Maghreb banking and telecom.
We're taking one insurance design partner in India and one in the US.
The first benchmark covers real-estate calls, which is where our production data came from. For insurance, we will build the first fine-tuned code-mixed model on live call archives — with legal review, custom vocabulary, and a public scorecard.