Frontier Models On Device
Palo Alto, CA
Joined November 2023
- Tweets349
- Following25
- Followers4.5K
- Likes940
We are hiring research interns who want to push the frontier of voice AI!
Check out our latest research here:
argmaxinc.com/blog/interspee…
Benchmarks
We benchmarked 6 speaker diarization systems across 11 datasets representing diverse and challenging real-world conditions using OpenBench: github.com/argmaxinc/OpenBen…
- Nemotron 3 Diarization achieved the lowest error rate (DER) across all 6 systems.
- It also scored the lowest across 5 of the 11 datasets despite its maximum speaker constraint of 8.
- This represents a huge leap in accuracy compared to its predecessor Sortformer v2.1 whose error rate was 60% higher.
Argmax Pro SDK 3.0.6 is in stable release!
Test real-time Qwen3-ASR on your iPhone or Mac without writing code at Argmax Playground: testflight.apple.com/join/Q1…
Introducing Argmax Pro SDK 3!
- Frontier* multilingual speech-to-text on iOS with Qwen3-ASR-1.7b
- Real-time API with custom vocabulary and word timestamps in under 2 GB
- New Dictation API optimized for post-recording latency
Also in this release:
- By popular demand: Unpredictable recompilation latency from Core ML is gone!
- Canonical formatting for dates and numbers in transcripts
TestFlight link in replies.
*Ranks at the top of the OpenASR private leaderboard. Private data deters leaderboard hacking.
Argmax Playground 3.0.2 is out on TestFlight with Qwen3-ASR-1.7b on iOS!
testflight.apple.com/join/Q1…
Introducing Argmax Pro SDK 3!
- Frontier* multilingual speech-to-text on iOS with Qwen3-ASR-1.7b
- Real-time API with custom vocabulary and word timestamps in under 2 GB
- New Dictation API optimized for post-recording latency
Also in this release:
- By popular demand: Unpredictable recompilation latency from Core ML is gone!
- Canonical formatting for dates and numbers in transcripts
TestFlight link in replies.
*Ranks at the top of the OpenASR private leaderboard. Private data deters leaderboard hacking.
Introducing Argmax Pro SDK 3!
- Frontier* multilingual speech-to-text on iOS with Qwen3-ASR-1.7b
- Real-time API with custom vocabulary and word timestamps in under 2 GB
- New Dictation API optimized for post-recording latency
Also in this release:
- By popular demand: Unpredictable recompilation latency from Core ML is gone!
- Canonical formatting for dates and numbers in transcripts
TestFlight link in replies.
*Ranks at the top of the OpenASR private leaderboard. Private data deters leaderboard hacking.
For current Argmax customers, this new model works with the APIs you are already familiar with!
Simply update your dependency to 3.0.2 and get started with the examples in the CHANGELOG: app.argmaxinc.com/docs/chang…
1
393
Apple restricted background access to the Neural Engine on iOS 27... and it's back!
This would have broken on-device inference for many apps that rely on this access, including many of our customers.
We made a strong case to Apple that this access should be reinstated, channeling our customers' voice.
We are elated that Apple reinstated this capability in the first iOS 27 Public Beta!
Here is the step-by-step guide in our docs: app.argmaxinc.com/docs/faq#q…
PSA: Argmax Pro SDK Kotlin 1.2.3 is now live with Tensor TPU support
Argmax TPU Engine is built with @googledevs LiteRT and Google Tensor ML SDK. Learn more about Google LiteRT here: developers.googleblog.com/bu…
1
5
1,583
Argmax OSS Swift hits 1.0!
- Swift 6 compatibility
- Now dependency-free to streamline Enterprise procurement, minimize conflicts and security risks
- Renames the package to argmax-oss-swift, consolidating WhisperKit, SpeakerKit and TTSKit
- Battle-tested in production on millions of devices for over 26 months
Release notes and upgrade guide: github.com/argmaxinc/argmax-…
344
Argmax OSS (formerly WhisperKit) just crossed 10M monthly on @huggingface!
- First ever Apple Silicon-only model to cross 10M
- Usage grew 10x in ~100 days
- Free, MIT Open-source and pure Swift
We are thrilled that WhisperKit reached 1 million monthly on @huggingface!
- First ever Apple Silicon-only model to reach 1M
- Usage grew 10x in 2025
- Free, MIT open-source and pure-Swift
On-device is becoming the default way to deploy commercial-grade speech recognition. Join the movement: github.com/argmaxinc/argmax-…
3
428
Google just published a blog post on the real-world commercial adoption of their new on-device inference runtime, LiteRT!
Heidi Health and Argmax are highlighted as the prime example of running medical transcription on Android devices, improving reliability, speed, and privacy while maintaining parity in accuracy with the cloud alternative.
WhisperKit is now Argmax OSS!
As part of our continued commitment to open-source, we are releasing part of Argmax Pro SDK, extending WhisperKit beyond speech-to-text.
Argmax OSS now includes:
- SpeakerKit: Add speaker info to your transcripts with the fastest implementation of Pyannote.
- WhisperKit: One of the most popular frameworks to deploy Whisper with 6+ million monthly downloads.
- TTSKit: Run Qwen3-TTS with real-time generation and playback for voice agents and content readers.
TTSKit
Argmax TTSKit is the newest addition to our family of on-device inference frameworks. The launch version is built around Qwen3-TTS 0.6b and 1.7b. Check out our open-source example app to get started: github.com/argmaxinc/argmax-…
In the short term, we are planning to implement voice cloning and optimize inference speed even further to enable real-time generation on iOS. Currently, real-time speed is only available for macOS.
Over time, we will keep adding leading text-to-speech models as they are released. If you have new model support requests, please file an issue on GitHub!
Argmax Playground for Android is now on Google Play!
Introducing Real-time Transcription with Nvidia Parakeet on Android!
Argmax Pro now supports Android with our brand-new Kotlin-first SDK, bringing Argmax's top-tier accuracy and real-time performance from Apple to Android.
Enjoying seamless NPU and GPU acceleration by @GoogleAI LiteRT across several major hardware vendors.
Links to blog and test app are in the replies.
Introducing Real-time Transcription with Nvidia Parakeet on Android!
Argmax Pro now supports Android with our brand-new Kotlin-first SDK, bringing Argmax's top-tier accuracy and real-time performance from Apple to Android.
Enjoying seamless NPU and GPU acceleration by @GoogleAI LiteRT across several major hardware vendors.
Links to blog and test app are in the replies.
Our blog post explains the technical breakthroughs and how Argmax became the first SDK to build with LiteRT in collaboration with @GoogleAI argmaxinc.com/blog/argmax-pr…
1
2
458
The latest Argmax Playground is now available on iOS TestFlight! testflight.apple.com/join/Q1…
TTSKit is now part of Argmax Open-source SDK!
Get started with the 0.16.0 release: github.com/argmaxinc/Whisper…
TTSKit now achieves sub-100ms time-to-first-byte for Qwen3-TTS 1.7b on Apple Silicon!
Link to the code repo and details in comments.
We are open-sourcing TTSKit!
Run state-of-the-art text-to-speech models on your Mac and iPhone.
The launch version supports @Alibaba_Qwen Qwen3-TTS and generates audio faster than real-time playback with sub-200 ms time-to-first-byte.
Voice cloning and advanced speed optimizations will be in the next version. Link to the GitHub repo and models on @huggingface in comments.
This is achieved by prompt prefix caching, and the framework and app example code is live on GitHub: github.com/argmaxinc/Whisper…
From Apple M2 through M5, the empirical range varies from 70ms to 130ms. TTSKit achieves faster-than-real-time playback on M2 and newer. We are working on further speed optimizations to achieve the same on M1.
The next feature is voice cloning!
