Stack Innovations
Start a project
Stack Innovations
Start a project
Real-time voice systems that hear a caller, transcribe every word, understand the intent, and respond out loud — fast enough to feel like a conversation. Streaming ASR, Claude for understanding, natural speech back, built for low latency and shipped to production.
A ledger of named systems where the line that moved was a call answered, a note captured, an intent resolved — handle time, accuracy, CSAT, accessibility. 6 of 30 shown · ledger updates as systems scale.
Press play and watch the loop run: audio streams in as a waveform, the transcript appears word-by-word, the intent is detected, a response is composed, and a voice speaks it back. Toggle diarization to label speakers; switch latency mode to trade accuracy for speed.
Quality hides in the stages between the microphone and the speaker — how you capture, where you transcribe, who decides the intent, how fast you speak back. This is the room we work in: each stage measured, each tool chosen for a reason.
Before a single byte of audio is captured, we map what callers actually say, what counts as a resolved intent, and where the latency budget bites. Then we build an evaluation set — the utterances we'll grade every release against.
Stream real audio — browser, phone, or field device — gate it with voice-activity detection, then transcribe with streaming ASR so partial words land before the caller finishes. Noisy audio transcribes to noise; this is where most voice systems quietly fail.
Attribute each segment to a speaker, then hand the transcript to Claude to classify intent, fill slots, and decide the next action. Hybrid from day one — fast classifiers for the common asks, Claude for the fuzzy ones.
Compose the reply — answer, confirm, or hand to a human — then stream it through text-to-speech so the voice starts before the sentence is finished. Barge-in lets the caller interrupt; an honest "let me get a person" beats a confident wrong answer.
Run the frozen eval set every change and score it — word error rate, intent resolution, p95 latency, abstain rate. No "sounded fine in the demo." A number that moves, or the change doesn't ship.
Live with caching for cost, guardrails for safety, and logging on every turn. We watch WER and latency as call volume grows, catch drift before callers do, and keep the system listening as accents, noise, and asks change underneath it.
Every eval run feeds the next: missed words sharpen the audio pipeline, failed intents tighten the prompts, slow turns justify the streaming work. Ship-and-forget voice drifts as accents, noise, and new asks pile up. Evaluated and tended, accuracy compounds.
The models, services, and tools we actually wire together to capture, transcribe, diarize, understand, and speak. No mystery framework — just the kit that keeps voice fast and accurate.
A free voice audit to start — bring a stack of your real call recordings and a list of real asks, and we'll show you what a streaming, speaking system would transcribe and resolve, and where today's IVR falls down. A prototype, not a pitch.
Get a voice audit →