60db community · voice AI builders in India and beyond
The voice AI community in India for people who ship voice agents.
A free WhatsApp group for founders and engineers running voice AI in production. Ask about latency, turn-taking, Hinglish speech-to-text and voice quality, and get answers from people who fixed the same thing last week.
- ✓ Straight answers on real call problems, from builders using every major stack
- ✓ First look at new 60db voices, models and pricing
- ✓ An optional 30-minute call with our team about what you're building
- builders
- 100+ builders
- countries
- 10+ countries
- leave anytime
- Free leave anytime
- 1. Your details
- 2. Intro call
- 3. Join
The problem
Voice agents break in ways a demo never shows you.
Your agent sounds great in the playground. Then it meets a real phone line, a real customer and a real accent, and things start to go wrong.
The pause that kills the call
A caller finishes speaking and waits. One second feels fine on your laptop. On a phone call it feels like the line dropped. People start repeating themselves, talk over the agent, or hang up. You know latency matters, but you can't tell which part of the chain is slow: speech-to-text, the language model, text-to-speech or the phone network in between.
The agent that won't stop talking
The caller says "no, wait" and the agent keeps reading its script. Or it stops at every cough and breath. Getting turn-taking and interruptions right is the difference between a call that feels human and one that feels like an old phone menu. Docs explain the settings. They don't tell you which values work for your callers.
The accent your model never heard
Your transcripts look perfect in English test clips. Then a caller mixes Hindi and English in one sentence, says a street name, or reads out an order number, and the transcript falls apart. Every wrong word sends the conversation further off track.
The bill nobody planned for
Per-minute pricing looks small until you add speech-to-text, the model, text-to-speech, telephony and retries. Many teams only learn their real cost per call after the first big invoice. By then, changing vendors means rewriting half the pipeline.
None of this is rare. Almost every team shipping voice AI hits the same walls in roughly the same order. The fastest way past them is to ask someone who already climbed over. That is what this group is for.
How it works
Three steps, about a minute.
Step 1 · Tell us who you are
Name, email, company, role, WhatsApp number and the voice tools you use today. It takes under a minute. We ask so we can welcome you properly and connect you with members who use the same stack or work on the same kind of product.
Step 2 · Book a 30-minute call (optional)
Pick a time that suits you and talk through your use case with the 60db team. We'll listen to where you're stuck and point you to the right people and resources. Rather just read and learn first? Skip it. You still get the join button.
Step 3 · Join on WhatsApp
One tap opens the group in WhatsApp. Say hello, share what you're building in a line or two, and ask your first question. There's no test, no fee and no waiting list.
Voice AI in India
Why India needs its own voice AI community.
Building voice agents for Indian callers comes with problems most global guides skip. This is where builders in India compare notes on them.
Hinglish and code-mixing
Callers in India switch between Hindi and English inside one sentence, and many switch into Tamil, Telugu, Marathi, Bengali or Kannada too. Speech-to-text models trained mostly on clean English often break on this. Members compare which models hold up on real mixed-language calls, and which ones still trip on names, addresses and numbers.
22 scheduled languages, not one
The Indian Constitution lists 22 languages in its Eighth Schedule, and real callers bring regional accents to each of them. A voice agent that works in Delhi can stumble in Chennai or Kolkata. The group is where builders share which voices and transcription models they tested for each language, and what they found.
Phone lines, noise and networks
A lot of voice AI in India runs over ordinary phone calls, from busy streets, shops and moving vehicles, on mobile networks that vary across the day. Audio that sounds clean in testing often arrives compressed and noisy. Members share how they test for this before launch.
Cost per minute in rupees
Many Indian products serve price-sensitive customers at high call volumes, so a few paise per minute decides whether a voice agent is worth running at all. Members compare real cost per call minute across speech-to-text, models, text-to-speech and telephony, and where they cut it.
Calling rules and consent
Commercial calls in India fall under TRAI's rules, including the Do Not Disturb registry, and personal data falls under the Digital Personal Data Protection Act, 2023. Members share how they handle consent, recording notices and outbound calling within those rules. It's peer experience, not legal advice.
Built by a team that works on Indian speech
60db builds speech and language models with Indian languages in mind. On our published benchmarks, Garuda ranks first for Hindi speech-to-text by word error rate. Our engineers are in the group to answer questions on Indian-language voice AI directly.
What's inside
The things docs don't tell you about voice AI.
Most voice agent problems only show up on real calls with real people. The group is where builders compare notes on them, openly and without a sales pitch.
Fix latency faster
See what other teams measured end to end, and which changes actually shortened the pause before their agent replied. Streaming speech, shorter prompts, smaller models, closer servers: people share what helped and by how much, so you can try the likely fix first instead of guessing.
Pick tools with real data
Ask people running ElevenLabs, Vapi, Retell, Deepgram, LiveKit, 60db and others what broke at scale and what didn't. Vendor pages tell you the best case. Members tell you what happened at two in the morning when traffic spiked.
Get early access
Try new 60db voices, models and features before they go public. Tell us what's missing, what sounds wrong and what you'd pay for. Members' feedback goes straight to the people building the models, and it shapes what ships next.
Find partners and hires
Meet founders, engineers and agencies working on the same problems in your market. Some members are looking for a co-builder, some for a contractor who has shipped a phone agent before, and some just want a second pair of eyes on an architecture.
Learn from other languages
Members build across more than ten countries. If you're shipping in Hindi, Tamil, Arabic, Spanish or a mix of languages in one call, someone has likely tested the same voices and transcription models and can tell you what held up.
Ask the 60db team directly
The people who build 60db's speech, language and memory models are in the group. Ask how something works, report a bug, or ask for a feature. You'll get an answer from an engineer, not a support script.
Topics
What members talk about.
The group stays on voice AI. Here are the subjects that come up again and again.
Speed
Time to first audio, streaming speech-to-text and text-to-speech, where to host models, and how to stop slow tool calls from freezing a conversation.
Turn-taking
End-of-speech detection, barge-in, backchannels like "mm-hmm", and how to let callers interrupt without the agent cutting itself off on every noise.
Telephony
SIP trunks, phone numbers in different countries, call recording, transfers to humans, voicemail detection and the audio quality you actually get over a phone line.
Voices
Choosing a voice that fits your brand, cloning with permission, pronunciation of names and numbers, and keeping one voice steady across a long call.
Testing
How to test an agent before real customers do: scripted test calls, scoring transcripts, catching made-up answers, and tracking quality after every prompt change.
Cost
Real cost per call minute, where money leaks, caching, choosing cheaper models for easy turns, and when it's worth switching providers.
Memory
Remembering a caller between calls, what to store and what to forget, and keeping memory lookups fast enough for live speech.
Compliance
Consent to record, telling callers they're speaking to AI, data residency and the rules that differ from one country to the next.
Going live
Rollout plans, fallbacks when a provider goes down, monitoring live calls and what to watch in the first week after launch.
Use cases
Whatever you're building with voice, someone here is too.
Members work on very different products. The problems underneath are often the same, which is why one person's fix so often helps another.
Customer support
Agents that answer inbound calls around the clock, look up orders and bookings, and hand over to a person when a caller is upset or the question is too hard. The hard parts: speed, knowing when to transfer, and handling callers who don't follow the script.
Sales and follow-up
Outbound agents that confirm appointments, qualify leads and remind people about payments. The hard parts: sounding natural in the first five seconds, voicemail detection, and the rules about calling people in each country.
Healthcare
Booking, reminders, intake questions and after-care check-ins. The hard parts: privacy, careful wording, older callers, and clear handover to staff for anything medical.
Recruiting
Screening calls that ask the same first questions to every candidate and pass a summary to the hiring team. The hard parts: fairness, accents, noisy rooms and keeping candidates comfortable talking to AI.
Content and dubbing
Voiceovers, audiobooks, e-learning narration and dubbing into other languages. The hard parts: pronunciation, emotion, keeping one voice steady across hours of audio, and the rights to use a voice.
Apps and games
Voice inside products: assistants, characters that talk back, and hands-free controls. The hard parts: speed on mobile networks, cost at scale, and voices that fit a character.
Day one
Questions you can ask the moment you join.
Not sure how to start? These are the kinds of questions the group is built for. Short, specific and tied to a real problem.
"Why does my agent pause before every answer?"
Share your stack and roughly how long the pause is. Members will help you work out which step is slow and what to change first.
"Which speech-to-text handles Hinglish best?"
People who tested several models on mixed-language calls can tell you what worked for them, and where each one still makes mistakes.
"How do I stop the agent talking over callers?"
Get settings and approaches others use for interruptions and end-of-speech detection, and what they changed after listening to real recordings.
"What does a phone agent really cost per minute?"
Compare your numbers with teams at a similar size, and find out which costs people forgot to include the first time.
Lessons
Mistakes the group helps you skip.
The same few mistakes cost teams weeks. Members spot them early, because most of them made them once too.
Testing only in quiet rooms
Agents tested on clean laptop audio often fail on real phone lines, in cars and on busy streets. Members share how they test with noisy, low-quality audio before launch, so the first bad call isn't from a paying customer.
Choosing a vendor on one number
A single benchmark rarely tells the whole story. Speed, accuracy in your language, cost, uptime and support all matter. Members tell you which trade-offs showed up later, after the contract was signed.
Writing prompts like chat
A good chat prompt can make a bad voice agent. Long answers, lists and links don't work when spoken aloud. Members share how they rewrote prompts for the ear: short sentences, one question at a time, and clear next steps.
Launching without a way out
Every voice agent meets a caller it can't help. Teams that plan the handover to a person, and a fallback when a provider goes down, keep their customers. Ask how others built theirs before you need it.
Who's in it
100+ builders across 10+ countries.
The group is small on purpose, so questions don't get buried. Members are people who build and run voice products every day.
Founders
People starting voice AI companies who need to make fast decisions on vendors, pricing and what to build first, without hiring a team to research every option.
Engineers
Developers and ML engineers wiring up speech, language models and telephony, who want answers on specific bugs and settings rather than general advice.
Product leads
People responsible for how a voice agent feels to customers: tone, speed, when to hand over to a human, and how to measure whether calls are going well.
Agencies
Teams building voice agents for clients across support, sales, healthcare, recruiting and more, who see many use cases and share what repeats.
Who runs it
Hosted by the team behind 60db.
60db builds speech and language models for voice AI. We started this group because the best fixes we've seen came from builders talking to each other, not from docs.
What we build
Kaak for multilingual text to speech, Garuda for multilingual speech to text, Vyas for language and reasoning, Smriti for memory, and Judge for decisions inside agents. Plus voice cloning, phone agents, telephony and an MCP server for Claude, ChatGPT and Cursor.
What we've measured
On our published benchmarks, Garuda ranks first for Hindi speech-to-text by word error rate. Kaak streams Hindi and English speech in under 200 ms. Smriti scores 92.4% on LongMemEval-s. You can read the full method for each on our benchmarks page.
Why we host it
We learn what to build next from people shipping real products. You get answers and early access. We get honest feedback. That trade is the whole business model of the group.
What we won't do
We won't turn the group into an ad channel. We share 60db news when it's useful, and we're glad to hear when another tool works better for your case.
See the method behind each number on our benchmarks page.
Your first week
What happens after you tap join.
No long onboarding. Here's what to expect in your first few days.
You get a welcome
Someone from the 60db team says hello and asks what you're building. Reply with a line or two: your product, your stack and the one thing that's bothering you most right now. That's your introduction done.
We make introductions
Using the role and tools you shared in the form, we point you to members working on something close to your problem. Talking to one person who has already shipped what you're building can save you weeks.
Your intro call, if you booked one
You get a calendar invite by email. On the call we go deeper on your use case, answer what we can on the spot, and follow up in the group with anything we promised to find out.
You ask your first question
Post it in the group whenever you're ready. Be specific about your setup and what you've tried. Then keep an eye out for the next person stuck on something you've already solved.
Why WhatsApp
Where you already are, not one more app.
We thought about Slack and Discord. We chose WhatsApp on purpose.
No new login
Most builders in India and our other markets already have WhatsApp open all day. Joining takes one tap, and you don't need another app, password or workspace.
Voice notes work
For a voice AI group, it helps to share how something sounds. Send a short clip of a bad call or a new voice and let people hear the problem for themselves.
You stay in control
Mute the group, archive it or leave it whenever you like. Catch up when it suits you. Nothing here needs an instant reply.
Is it for you?
Built for people shipping, not browsing.
A great fit if you
- ✓ Build or run voice agents, IVR, dubbing or text-to-speech products
- ✓ Have a prototype, a pilot or real calls in production
- ✓ Use any stack, 60db or not
- ✓ Like sharing what worked and what didn't
- ✓ Want straight answers from people who have done it
Probably not for you if you
- ✕ Want a job board or a reseller program
- ✕ Plan to post promotions and cold pitches
- ✕ Are only curious about AI in general
- ✕ Expect free consulting on demand
- ✕ Don't want to share anything back
House rules
How we keep it worth reading.
A group is only useful if the signal stays high. These few rules keep it that way.
Stay on voice AI
Questions, findings, tools and launches related to voice. General AI news and memes belong somewhere else.
No cold pitches
Don't message members to sell to them. If someone asks for a recommendation, you're welcome to answer it honestly.
Share back
If the group helped you fix something, post what worked. Your answer will help the next person who hits the same wall.
Be specific
Include your stack, numbers and what you already tried. Specific questions get specific answers, usually faster.
Respect privacy
Don't share customer data, call recordings with personal details, or anything you're not allowed to share.
Be kind
Everyone here was new to voice AI once. Disagree with ideas, not people. Repeat offenders get removed.
Why join
Figuring it out alone vs. with the group.
| Situation | On your own | 60db community |
|---|---|---|
| When a call sounds wrong | Search forums and old GitHub issues | Ask people who hit the same issue |
| Choosing a vendor | Marketing pages and benchmarks | Notes from teams running it in production |
| Fixing latency | Trial and error, one change at a time | Start with what worked for others |
| Indian languages | Test every model yourself | Learn from members who already did |
| New models and voices | Hear about them after launch | Try them early and give feedback |
| Talking to model builders | Support tickets | Ask the 60db engineers directly |
| Cost | Your time | Free |
Your details
What we do with what you share.
We ask for six things. Here is exactly why, and what happens to them.
Why we ask
Your name and company help us welcome you. Your email is where we send your call invite and the occasional important update, like early access openings. Your role and tools help us introduce you to members with the same stack or problem. Your WhatsApp number is how the group works, and it tells us your time zone so we can suggest call times that suit you.
Where it goes
Your details are stored securely by 60db and used only to run the community and, if you book one, your intro call. We don't sell them, and we don't hand them to other companies.
What we won't do
We won't cold-call you, add you to sales email sequences or share your number outside the group. Inside WhatsApp, other members can see your number, as in any WhatsApp group.
If you change your mind
Leave the group anytime from WhatsApp. Want your details deleted too? Email support@60db.ai and we'll remove them.
Questions
Before you join.
Free · about a minute
Stop debugging voice agents alone.
Join India's voice AI community: 100+ builders across 10+ countries who share what works on real calls. Fill in six fields, book a call if you want one, and you're in.
Join the community →