Lower your voice AI cost per connected minute.
On your own calls, I compare three options: tuning what you run now, a cheaper API, or quantized open models in your own infrastructure. I recommend whichever wins, even if that's a vendor API.
For voice-agent teams with live call volume and a voice bill to lower. In 10 business days, I audit one costly component in one language, using your last 30 days of invoices and a call sample tested in your environment.
The Voice Cost Audit
One fixed-price audit of one part of your voice stack, usually the one that costs you most per minute, on one language. Call samples are analysed in your infrastructure and I never take a copy. API alternatives are first priced from your usage and vendor rates; any live API test happens only if you approve it from your own account.
Your baseline measured; alternatives priced and checked
Your cost per minute split by speech-to-text, the LLM turn, text-to-speech and telephony. Concurrency, latency and error rate measured on a sample of your own calls, inside your infrastructure; I never take a copy. The cheaper-API option is priced from vendor list prices and the call profile I measure, so no audio is sent anywhere. Then a written plan: what to change, what each step costs, and the estimated savings, with the assumptions and checks still required.
$1,000 fixed, or ₹75,000 for India-based teams
Full refund if the audit finds possible savings under 20% of your current all-in cost per connected minute, from your last 30 days of invoices. Every option counts toward that 20%, including a vendor API. If self-hosting wins and you run a pilot with ILLATE, half the audit fee is credited to it.
If you pay under 6 cents a minute and run under 100,000 minutes a month, I'll tell you before you pay that the savings may be small, so you can decide if it's worth it.
How it works
Scope
A 20-minute call, no data needed: how your stack is built, minutes a month, languages and your latency target. Then we pick one costly voice component and one language.
→ a one-page scopeBaseline
I read 30 days of invoices and measure a representative call sample inside your systems.
→ your baseline cost per connected minuteThree paths
I price the three paths on the same workload and compare quality, latency, reliability and cost.
Recommendation
You get the evidence table, the recommendation and the refund decision within 10 business days.
→ one written recommendation
A possible next step, separate from the audit: if self-hosting wins, we can run a pilot of about six weeks, and half the audit fee is credited to it. I deploy the optimized serving in your infrastructure. It's accepted only if the error rate stays within a margin we agree of your current model on held-out calls, latency meets your target, and the saving is measured. Pricing comes from the measured savings, agreed in writing before it starts. If tuning your stack or an API wins, you get the plan and you're done.
Want to see the arithmetic first? A worked example: one voice stack, priced three ways, from public list prices. It's an example, not a client result. There's also a synthetic worked audit, not a client result, with speech-to-text measured on one CPU machine.
What I can show you today
- Background. Ternary and low-bit LLM quantization and GPU kernels at OneBit AI, and fixed-point inference at Trebuchet Solutions.
- How I test. A benchmark with its rules written down before the run, paired confidence intervals and every number published, including where my own model didn't win (the Banking77 benchmark, the open-source parity tool). It's text classification, not speech, but the testing discipline is the same.
- Next. On the first call I'll show you exactly how I'd measure your cost per minute.
Public test sets aren't your calls. Your accents, noise and call mix will differ, which is why the audit measures yours, on your servers.
What I don't do
I don't build voice agents
I don't sell a voice agent platform and won't compete with you. I make the models underneath yours cheaper to run.
I don't need your recordings
Not to start. For the audit I work on them inside your infrastructure and never take a copy. If you want an API tested live, you run it from your own account, so audio goes only to a vendor you choose. Nothing I promise rests on numbers I haven't measured on your calls.
Poojith Devan, inference engineer, New Delhi
The person on your first call does the measuring and the optimization. ILLATE is one engineer, so I work with a few voice teams at a time.
Questions
Do my recordings leave your infrastructure?
No, not for the audit. I analyse them where they are. Audio only reaches an API vendor if you choose to test one live from your own account.
We use a speech API, not our own models. Is this for us?
Yes, it's built for API users too. The audit prices three options on your calls: tuning what you run now (prompt caching, voice choice, silence billing), switching to a cheaper API, and self-hosting quantized open models. If staying on an API wins, that's what I recommend, and nothing more is owed.
Which languages?
English is well covered by open speech models; other languages vary, especially for text-to-speech. For Indian languages, open speech-to-text covers all 22 scheduled languages and open text-to-speech covers fewer, so we check yours on the first call. The audit starts with one language.
What I work on
Open speech models
Open speech-to-text such as Whisper (MIT / Apache-2.0) and NVIDIA's Parakeet (CC-BY-4.0), open text-to-speech such as Kokoro (Apache-2.0), and small open LLMs. For Indian languages, AI4Bharat's IndicConformer for speech-to-text, IndicF5 and Indic Parler-TTS for text-to-speech, and Sarvam's open LLMs. If you use a speech API or OpenAI's speech models, the audit compares it with a cheaper API and an optimized open model.
More calls per GPU, or no GPU
INT8 and lower-bit quantization, fixed-point inference on CPUs, and batching and streaming tuned for concurrent calls. Quality is checked at every step against your current model, not assumed. When you self-host, no call audio goes to a model vendor.
API price baselines the audit compares against: Mistral's Voxtral 3, Grok Voice 3.0 and ElevenLabs' Eleven Realtime. These are vendor APIs, priced at the rate on the day of your audit, not models I self-host.
Parakeet: NVIDIA's parakeet-tdt-0.6b models, licensed under CC BY 4.0.
Other work: OpenAI fine-tune migrations are a separate offer.
Get a 20-minute cost estimate
Email me your minutes a month, your languages, and which speech-to-text, text-to-speech and LLM you use. No data needed. I reply within one business day with times for a call.