... speech-to-text part is already streaming and feels fast . But then... The user stops speaking transcript is ready ...nothing happens for a moment AI finally starts answering. That gap is surprisingly noticeable. My current pipeline is roughly: Microphone Deepgram Streaming STT detect question LLM streaming response And now I'm trying to understand where those extra milliseconds are actually going . There are potentially several places: * Deepgram

... built for that moment. winInterview is a desktop app that floats as an invisible overlay on Zoom/Meet/Teams and assists you in real time during a job interview. It listens to the audio (microphone + system) and transcribes it with Deepgram; an LLM (DeepSeek) suggests answers instantly, and with Ctrl+Shift+C/A it can read screenshots of your screen to solve technical exercises. It's a paid app with plans (free/lite/pro/ultra). The higher-tier plans unlock invisible mode, multiple screenshots, and interview ...

... punctuated sentence you meant. A few things that make it different: On-device or cloud, your choice on-device means your audio never leaves the phone. Meeting bot that joins Zoom/Meet/Teams for a transcript + drafted follow-up. $9.99/mo vs Wispr Flow's $15 , cloud-only, no meetings. Free tier, no card. BYOK : Bring your own API keys Use your own Deepgram

... punctuated sentence you meant. A few things that make it different: On-device or cloud, your choice on-device means your audio never leaves the phone. Meeting bot that joins Zoom/Meet/Teams for a transcript + drafted follow-up. $9.99/mo vs Wispr Flow's $15 , cloud-only, no meetings. Free tier, no card. BYOK : Bring your own API keys Use your own Deepgram

... upfront. GPT-4 tiles a 1024 1024 image into ~765 tokens before it reads your prompt. Send 10 images in a conversation, and your "cheap" chat call isn't cheap anymore. GPT-4o Realtime charges $100/1M audio input tokens vs $5/1M for text. Same model. 20 difference depending on how you send data. Batch API discounts give you 50% off on OpenAI, Anthropic, and Google but only if your architecture is deliberately designed around it. Most aren't. Agentic loops ... ... that trigger web searches can quietly run up to $10 $35 per 1,000 calls, depending on the provider. One pipeline can fire dozens. STT providers round audio duration differently. AWS rounds up to 15 seconds. Deepgram

... rivra_dev  Ha, thanks. Honest answer: I'm cheating on diarization. Discord's voice gateway hands you per-user PCM streams via voice_recv, each one tagged with the speaker's user_id. discli just opens a separate Deepgram (or Whisper) session per speaker, so three people talking at once isn't a cocktail party problem, it's three parallel websockets. Transcript events come out interleaved by timestamp with user_id attached. Where it breaks: a hot mic picking ...

... responds. That's why every voice agent feels slow and robotic. Our model works differently. It listens to the rhythm and intonation of speech and predicts the end of turn in milliseconds. We tested it against SmartTurn, LiveKit, and Deepgram Flux. Full benchmark table and our public test dataset are in the technical Technical blog link: /blog/turn-taking-v3 Comment from Alper Tayfur(@alpertayfurr): This is a very real production problem for voice agents. The jump from silence-based turn detection ... ... rhythm/context-aware turn prediction feels especially important, because most awkward AI calls break when the agent interrupts too early or waits too long. The mhm vs

... most excited about. OpenCut-AI is built with a pluggable engine architecture. Adding a new voice or transcription provider is straightforward, and we want the community to drive what comes next. Models we're considering: - ElevenLabs (premium voice synthesis) - Deepgram (real-time STT) - Fish Speech (open-source voice cloning) - Kokoro TTS (lightweight and fast) - StyleTTS 2 (human-level quality) Have a model you'd love to see? Drop it in the comments with: - What it does - Why it matters ...

... handle information. Comment from Farhad Asbaghipour(@farhad_asbaghipour): I d probably drop immediately trust is gone at that point. It s interesting how many modern workflows are becoming record first, ask later. Curious how others balance efficiency vs consent in client communication? Comment from Farhad Asbaghipour(@farhad_asbaghipour): @rohanrecommends  That s a really strong point once trust shifts, everything becomes transactional. I ve actually started seeing clients hesitate more, even before calls now. Feels like people are becoming more ... ... move to an international signal (a ping/beep or something) that tells people it's been recorded. I believe you need to know. Comment from Paul Veth(@paulveth): @rohanrecommends  true that! I can give this idea to my partner @Deepgram

... folks We've been shipping a steady stream of improvements to Layercode, based on your feedback. Here's what's new: Deepgram Flux improvements Fixed the 60-second silence disconnect bug that was interrupting some longer pauses. Deepgram

... excited to announce a major upgrade to Layercode: support for Deepgram Flux: the world's first transcription system built specifically for voice agents. What is Deepgram

... custom voices in Layercode. Sonic-3 + Flux on the edge = the fastest voice AI agents Sonic-3 is ideal for developers building conversational experiences where responsiveness and human-level expressiveness matter. We recommend pairing Sonic-3 with Deepgram Flux running on Layercode's edge network to give your Layercode voice agents the benefits of the fastest available STT + TTS combination: ultra-low latency, intelligent turn-taking, and near-human expressiveness delivered reliably to users anywhere in the world. Try Sonic ... ... under your agent settings. New to Layercode? Sign up for a developer account and get $100 in credits to test Sonic-3. Comment from Aidan Hornsby(@aidanhornsby): Thanks @fmerian ! We're excited about this one, especially when combined with Deepgram

... With these two developer-focused voice app launches today, we should have a bakeoff to see which one rules...! https:// ...

... bots that can answer calls, take orders, or redirect to a live agent. Train your assistant with your menu, FAQs, or custom logic. Plug in your favorite STT (speech-to-text) and TTS (text-to-speech) providers (like ElevenLabs, Deepgram, etc.) Launch on Twilio, SIP, or your phone number. Why it s different: No code needed. Realistic voice conversations, not robotic scripts. Fully customizable logic and context memory. We built Monobot because we were tired of clunky IVR menus ...

... month-long Voice AI course and community events series that starts this week. Join Voice AI experts from OpenAI, Google, NVIDIA, AWS for Zoom sessions (or watch them async). $10k in credits from Baseten, Cartesia, Cerebrium, Coval, Daily, Deepgram, fal, Fireworks AI, Freeplay, Gladia, Layercode, mem0, Modal, OpenPipe, PlayAI, Rime Labs, Tavus, Ultravox, Vapi, and Weights & Biases. Special Product Hunt deal through this link ($200 for the course, discounted from the $500 full price): https://maven.com/pipecat/voice-... Rajiv Ayyangar ...

... browsers are showing documentation, for example. I'm thinking of the screen as the "universal API" for getting as much context about the development environment as possible. Comment from Sharon Yeh(@_sharonyeh_): Hey @kwindla ! I m a PM at Deepgram, and my team is building a voice OS for developers. What you re describing aligns closely with our vision! Would love to chat further and / or keep you updated on our progress :) Comment from Kwindla Kramer(@kwindla): @rajiv_ayyangar ...

... making money in a way that's more aligned with connecting makers and customers. To be clear, I like our ad model - our advertisers are awesome (they're some of the biggest success stories on Product Hunt, like Notion, Deepgram, Intercom...) but I think there's a compelling direction to explore that's more like affiliate revenue. Comment from Rajiv Ayyangar(@rajiv_ayyangar): @nico_jeannen  we feature products that we believe our community should know about. Put another ... ... think is the single most important trait of successful launches on Product Hunt? How do you make it to the top of the upvotes? Comment from Corgi Cabal(@corgicabal): AI landscape questions: How much do you personally use google vs

... soon. Comment from Ankit Gordhandas(@agordhandas): @ankitgordhandas   @rajiv_ayyangar  even better! If the person on the other end starts responding in Spanish, it switches to Spanish. Comment from Nour Almasri(@almasri_no786): Just gave Deepgram a spin and wpw Feels like the voice AI equivalent of having super-hearing. Transcription is crazy fast and impressively sharp even picked up my half-mumbled test phrases Loving how dev-friendly it is, but now I m wondering how well does ...

... Daniel Mishler(@daniel_mishler): Late to the party, but I'd be curious of what you think of Vibe Coder (launched today, full disclosure: I built it using Cursor for exactly this use case). Open-source extension powered by @Deepgram , integrated into the IDE to make it easy. My favorite thing is the integrated voice agent, where you can talk to it naturally when you're just starting or brainstorming, then when you've got a good plan discussed ...