Abhyuday Patel

Product @voicearena_ai | LLM Inferencing | Voice Agents

[Shobhit Banga @shobhitbanga]

Drop 4/15: Announcing the Voice Arena Dictation Bench - the first open benchmark of AI dictation apps.

Not their APIs: the real apps. We played 761 recordings from 87 speakers into each of 10 apps on a Mac, then 5 LLM judges asked if each draft could be sent as-is.

Wispr Flow: 79%. Apple: 35% 🧵

[Voice Arena @voicearena_ai]

ElevenLabs launches Scribe v3 Live with top real-time speech-to-text accuracy 🚀

We tested it on 70+ hours of real, unscripted phone calls.

The new #1 streaming model in Romanian and Indian English, and first on all 10 metric splits we report there 🧵

ElevenLabs' own numbers:
• Scribe v3: 27% fewer word errors than Scribe v2
• Scribe v3 Live: 43% fewer errors than Scribe v2 Realtime

They benchmarked on read speech. We test on spontaneous conversation, where models usually struggle most.

Here's what we found 👇

Congrats @ElevenLabs 👏

[Abhyuday Patel @ogabhyuday]

@mustafasuleyman https://t.co/b8AtlVVgbS

[Voice Arena @voicearena_ai]

🏆 Rumble-2 from @reson8_official debuts at #1 on the Voice Arena ASR Leaderboard for Brazilian Portuguese.

Every model is tested on the same private set of 3,814 conversational clips. Rumble-2 ranks at the top with 9.84% Normalized WER, ahead of all 37 other models.

Reson8 now holds the top two spots, with Resonant-1 at #2 (9.85%).

Congratulations to the @reson8_official team! 👏🏻

[Abhyuday Patel @ogabhyuday]

@jrbluth @LererHippeau @cohere @twelvelabs @adaption_ai @WindBorneWx @DetectIfAI do check this for % of ai?

[Abhyuday Patel @ogabhyuday]

@claudeai @DetectIfAI does it have ai voice? if yes by how much?

[Detectif.ai @DetectIfAI]

introducing the DetectIfAI bot for X!

tag @DetectIfAI under a post with audio or video.

it'll check the audio and reply

- no AI detected
-AI detected
-partly AI, inconclusive.

[Tarini Sai Padmanabhuni @tarini_axory]

heard a voice on here and thought "there's no way that's real"?

we built this for that moment.

tag @detectifai under the post and it'll check the voice for you.

try it out and tell me what you think!

[Shobhit Banga @shobhitbanga]

Drop 3/15: Introducing the Speech Recognition Benchmark on @voicearena_ai.

One benchmark for every use case of speech recognition, in every language. Tested on real conversations.

Dataset 1 is live: the Monsoon Conversational ASR Test Set, two-person conversations in 5 languages.

Leaderboard and public test sets at Voice Arena.

[Rikiya Takehi @rikiyatakehi]

We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.

We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: https://t.co/bOWCcFhCVq

[Voice Arena @voicearena_ai]

🏆 NVIDIA Nemotron 3.5 Transcribe debuts at #1 among 14 systems on Voice Arena’s Diarization + ASR leaderboard.
Across the full benchmark, it achieves 32.43% cpWER.

This track tests whether a system can get both the words and the speakers right. We evaluated 139 held-out conversations with 2-8 speakers across in-person and online settings, including overlapping speech.

Nemotron 3.5 Transcribe is first in both settings: 34.08% cpWER in person (far field) and 22.13% on online calls.

It has the lowest word deletion rate on the leaderboard. Deletions are 8.32% of reference words, against a median near 31%.

Congratulations to @NVIDIAAI on the release ! 👏

[Abhyuday Patel @ogabhyuday]

@elonmusk This feeling right here is probably the reason I love tech so much.

[Abhyuday Patel @ogabhyuday]

@helloiamleonie sept 26' is crazy, hard to keep it up with news, can't wait to try on.

[Abhyuday Patel @ogabhyuday]

@polynoamial @dwarkesh_sp @OpenAI @kevinleestone @mikegmalek @__eknight__ @amuellerml @zhangir_azerbay @CheukHeiChu lfg ❤️‍🔥

[Abhyuday Patel @ogabhyuday]

Xiaomi is now publicly streaming a ~$1M+ agentic RL run.
I built a live forensic layer that watches its 18 telemetry signals and asks:
What changed? What caused it? What does it imply?
correlated step by step.
→ https://t.co/UbcEQCUF1Z

@_LuoFuli

[Abhyuday Patel @ogabhyuday]

Xiaomi is now publicly streaming a ~$1M+ agentic RL run.
I built a live forensic layer that watches its 18 telemetry signals and asks:
What changed? What caused it? What does it imply?
correlated step by step.
→ https://t.co/UbcEQCUF1Z

@_LuoFuli

[Abhyuday Patel @ogabhyuday]

@get_soju @localhostIND @localhosthq Super cool

[Harshit @get_soju]

"I think this version looks good enough"
"I think this one is going to be the final one"
That's me past 2 weeks....

Final look of Dokidek is ready.

Waitlist at https://t.co/ScS1Tc7DtH

@localhostIND @localhosthq

[Abhyuday Patel @ogabhyuday]

@boardyai @andrewdsouza building Orbitrage, multi-agent infrastructure for production routing, cost control, and failure attribution. want in on Boardy Pro

[Abhyuday Patel @ogabhyuday]

The winning move isn’t defaulting to the flagship. It’s intelligent routing. we shipped day-0 support for Fable 5 the moment it dropped. (https://t.co/ujBltPJbLw) If you’re already playing with Fable 5, try routing it intelligently with Orbitrage for free.

[Abhyuday Patel @ogabhyuday]

Fable 5 is here - @AnthropicAI just dropped their strongest general model yet.
Here are results of past 3 hours of deep testings -

[Abhyuday Patel @ogabhyuday]

It brings noticeably better judgment, coherence, and persistence over long contexts. Also its real pain with security/vulnerable tasks, its classifier auto routes to opus. But for the other 95% of everyday work? Opus (and even smaller models) are still smarter to use.