Sunday roundup: debugging habits and multilingual speech

One post this week, and it was a good one. The daily writing cadence is just getting started so most days are still empty, but Saturday landed solidly. Here is what I wrote and what else was happening.

Saturday: 5 habits that reduce voice AI debugging

The first Saturday post and the start of more regular writing. I covered five habits that cut voice AI debugging time from multi-hour investigations to five-minute checks. Log the raw API response before the SDK parses it. Save every audio file the pipeline touches. Isolate network latency from processing latency. Version your prompt and configuration changes. Test with the simplest possible input first.

That last one is the one I keep coming back to. A two-second WAV of me saying “hello” has caught more bugs than any integration test I ever wrote. Environment variable issues, encoding mismatches, stale dependencies. The simple test catches them all before they become clever.

The rest of the week

The other days were quiet on the writing front, but not on the voice AI front. A few things happened that set the tone for the weeks ahead.

Flux went multilingual. Deepgram rolled out Flux Multilingual on April 29, adding 10 languages with real-time language detection and mid-call switching. One model, one API, no routing logic. The big news here is code-switching within a single conversation, which is the kind of thing that looks like a nice feature until you build a voice agent for a multilingual market and realise how much complexity it removes.

AssemblyAI launched their Voice Agent API on April 29. A single WebSocket with STT, LLM orchestration, and TTS behind one connection. Live configuration updates without reconnecting. Tool calling with JSON Schema. Session resumption on dropped connections. $4.50 an hour flat. The pitch is that most teams spend weeks stitching together separate providers, and this collapses the whole pipeline into one integration.

Twilio updated Conversation Relay on April 22 with Deepgram Flux support, PCI and HIPAA compliance, and a drag-and-drop Studio widget for non-code workflows. The Flux integration matters because Twilio handles a huge share of telephony traffic, and better STT on that channel means fewer transcription errors in production voice agents that run over phone lines.

xAI also launched Grok STT and TTS APIs in mid-April, priced aggressively at $0.10 per hour for batch transcription and $4.20 per million characters for TTS. The benchmarks they published showed strong performance on phone call entity recognition, which is the hardest practical use case for speech-to-text. Five voices, 25 languages for STT, 20 for TTS. Another competitor in a market that keeps getting more interesting.

Microsoft made Copilot Studio real-time voice agents generally available in Dynamics 365 Contact Center on April 27. Enterprise-grade voice agents with turn detection, interruption handling, and the control plane that Fortune 500 procurement teams expect. The interesting thing here is that every major platform now has a voice agent offering. The question is shifting from “can I build a voice agent” to “which platform gives me the best developer experience.”

What I am still thinking about

The daily cadence needs to actually be daily. This week showed one post and a lot of catch-up. Next week I want to fill more days. The Saturday post proved the format works, so the bottleneck is just sitting down and writing.

The voice agent API landscape is moving fast. Four companies shipped significant voice AI products in the last two weeks of April alone. That pace is going to make it harder for developers to evaluate platforms objectively because the surface area keeps expanding. The ones that focus on developer experience will win the repeat integrations.

Frequently asked questions

Where is the Saturday debugging habits post?

At https://lukeocodes.dev/5-habits-voice-ai-debugging. Covers five habits that reduce debugging time for voice AI pipelines.

What languages does Flux Multilingual support?

English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian, and Dutch. All in a single model with real-time language detection and mid-call switching.

What does the AssemblyAI Voice Agent API cost?

$4.50 per hour flat for the full pipeline. STT, LLM reasoning, and TTS behind one WebSocket connection. No per-call metering or separate billing per provider.

How did Twilio add Flux support to Conversation Relay?

Developers using Conversation Relay can now use Deepgram Flux as their STT provider. This improves transcription accuracy and turn detection for voice agents running over Twilio’s telephony infrastructure.

Why start a daily writing cadence in May 2026?

I had been writing monthly posts on the first Tuesday of each month. The daily cadence is an experiment in consistency. Posting more often means covering more of what is happening in voice AI as it happens, rather than compressing a month of changes into one post.