
DevRel @Speechmatics. I build voice agents and speech demos for real-world audio. Code, AI, benchmarks, and the occasional talking game. Start building below.
London, England
Joined October 2012
- Tweets206
- Following298
- Followers173
- Likes136
Building voice AI? I work in DevRel at @Speechmatics.
I share runnable demos, integration notes and tests with real-world audio from voice agents to talking games.
Start here: github.com/speechmatics/spee…
What are you building, and where does speech fail?
Replying to @DropHammerai
Generally speaking, people were blown away that we were running our models on #NVIDIA Jetson. Data residency, GDPR, on-premises deployment, and air-gapped environments came up repeatedly
1
21
Edgars Adamovics retweeted
That's a wrap on #VapiCon! 🎤
From rapping to impressions, you put our On-Device voice AI to the test, running completely offline on @nvidia Jetson.
Thanks to the @Vapi_AI team for another brilliant year, and to everyone who stopped by!
Curious about On-Device speech recognition? Let's talk.
speechmatics.com/speech-to-t…
Replying to @voelker_thomas
From the conversations I had, it is rarely just one thing. Background noise is a challenge, but overlapping speakers and turn-taking came up even more often.
Names, numbers, and industry-specific terminology are another story, especially in healthcare and multilingual environments where accuracy really matters.
What is interesting is that developers are increasingly testing models on their own real-world audio rather than relying on generic benchmarks. That is where the differences really start to show.
1
1
18
Day 2 of #VapiCon is a wrap!
Two days of incredible conversations with developers, founders, and even competitors.
Big takeaways:
Multilingual AI goes beyond language support. Accents, dialects, and code-switching matter.
Real-world audio is still messy. Overlapping speakers, background noise, and turn-taking remain tough challenges.
Speech-to-speech vs. cascaded architectures? The debate is heating up!
And on-device AI is getting serious attention, especially around privacy and control.
We also wrapped up with TWO winners at the @Speechmatics booth!
One took home an RTX 5070 from our singing challenge, while a young champion smashed our voice-imitation leaderboard and won a LEGO Technic set!
My biggest takeaway? Voice AI isn't just about building better models. It's about the developers, the community, and what we build together.
Until next time, VapiCon!
#VoiceAI #Speechmatics
You got what I need! This was epic transcription! Love the energy today
This quoted post is unavailable.
2
40
Think speech-to-text can handle RAP?
Fast delivery. Slang. Words flying at ridiculous speeds.
Most speech recognition models struggle to keep up with normal conversations, let alone someone dropping bars.
So naturally, we turned the @Speechmatics booth at #VapiCon into a rap battle.
And put our models to the test LIVE.
Turns out, our STT has bars too. 🔥
Whos brave enough to challenge it next? There is still time!
1
66
Turns out, speech-to-text has a hidden boss level: KARAOKE.
At #VapiCon, @HelenZeng312453 put @Speechmatics to the test by singing Reflection from Disney's Mulan.
And it actually understood her singing!
Most speech recognition models are built for speech, not singing. Stretch the vowels, change the pitch, throw in some melody, and things get interesting.
Apparently, Speechmatics can handle Disney era too.
Who said STT can't do karaoke?
Replying to @Speechmatics
@Speechmatics made karaoke at VapiCon possible 👏👏 they detected lyrics from my singing perfectly 💯
3
116
Day 1 of #VapiCon is a wrap!
Spent the day talking with developers and founders building everything from restaurant ordering agents to healthcare, recruitment and construction AI.
A few things stood out:
Accuracy in real-world conditions is still a huge challenge. Turn-taking and overlapping speech remain difficult. And more developers want control over their voice AI stack.
We are also seeing conversations shift towards real benchmarks: P50/P95 latency, multilingual accuracy, diarization and testing on actual customer audio.
My biggest takeaway? The next leap in Voice AI isn't just about making agents talk better. It's about making them listen better.
Great conversations, brilliant builders and plenty to explore on Day 2!
#VoiceAI #Speechmatics
#Vapicon WIN from @Speechmatics Geforce RTX 5070. But how?
The booths ready. Looking forward to seeing what everyones building.
I am at #VapiCon at Fort Mason with @Speechmatics today. Bring your demo, come talk to me, tell me what you are stuck on, and come say hello. And win some prizes!!
Great question! This was actually one of the challenges we discussed.
Diarization can tell you multiple people are speaking, but that doesn't necessarily tell the agent who it should listen to.
One approach discussed was running speaker identification in parallel using a subagent, so the main agent isn't blocked while it determines who is speaking.
The tricky part is reliably distinguishing the intended caller from background voices without adding latency. Still a big challenge in real world environments!
2
18
Exactly! That is the key challenge with self-improving agents. Fixing one failure means very little if you introduce regressions elsewhere.
Ideally every proposed change should be tested against a separate, representative set of calls covering different accents, noise conditions and edge cases, with primary metrics improving and secondary metrics staying stable.
1
33
Your Voice AI agent works perfectly in a demo.
Then a real customer calls... and everything breaks.
Great discussion at our latest Voice AI meetup with @Speechmatics and @roark_ai !
We covered some big topics:
- Why real world testing matters more than clean demos
- Speech-to-speech vs cascaded architectures
- Speaker identification and real-time challenges
- Self improving agents that test, measure, and deploy their own improvements
My biggest takeaway? Building voice agents is getting easier. Making them reliable in production is where the real engineering begins.
Huge thanks to James and Daniel from Roark for flying over from SF and sharing their insights!
And yes, we had pizza, beer, and fireworks!
#VoiceAI #AIAgents
Replying to @MiaParker522 @Speechmatics
That's a good one. We have speaker retain capability; no other speaker can override the main speaker, but all speakers' info is retained. Certainly, come around and let's take it for a spin!
11
Deepgram acquiring Daily, the team behind Pipecat, is a big move for #voiceAI.
My take from @Speechmatics DevRel: developers win when they can mix models, benchmark on real conversations, and switch providers easily.
Keeping Pipecat open source and retaining its provider neutrality will matter in the long run, the time will tell.
1
10
306



















