AI & ML interests

None defined yet.

Recent Activity

aufklarer  updated a model 9 days ago
soniqo/VAP-OTO-ONNX
aufklarer  updated a collection 10 days ago
ONNX
View all activity

aufklarer 
posted an update 9 days ago
view post
Post
3,237
We’ve published Soniqo’s voice-agent leaderboard and an article explaining what we measured.

aufklarer/full-duplex-leaderboard

The board covers three parts of a speech system:

- Turn-taking · TurnBench: detecting turn endings, false alarms and decision delay.
- Commands · SLURP: choosing the correct tool and matching its argument values.
- Voice style · InstructTTSEval: following voice descriptions, delivery directions and role-play instructions, with separate scores for each.

Models include VAP, p99, Smart Turn, Qwen3-TTS, VoxCPM2, FireRedTTS3 and MOSS-VoiceGenerator, alongside cloud configurations from OpenAI, Google, AWS and ElevenLabs. Each table shows run sizes and links to the protocol.

We also published aufklarer/P99-Turn-1-Mini-CoreML, aufklarer/VAP-OTO-CoreML and soniqo/VAP-OTO-ONNX exports. Both models have native Swift APIs; VAP also has C++ and Android support.

Explore the leaderboard aufklarer/full-duplex-leaderboard · Read the article https://soniqo.audio/blog/voice-agent-benchmarks
  • 5 replies
·
aufklarer 
updated a model about 1 month ago
aufklarer 
posted an update about 1 month ago
view post
Post
239
⚡ GLiNER2.5-Decide now runs natively on Apple Silicon, in MLX Swift

Fastino's 340M decision model scores every label you give it, or finds entity spans, in a single forward pass. Nothing is generated. I ported it to speech-swift and published three MLX conversions:

- aufklarer/GLiNER2.5-Decide-340M-MLX-8bit (default)
- aufklarer/GLiNER2.5-Decide-340M-MLX-fp16
- aufklarer/GLiNER2.5-Decide-340M-MLX

Measured on an idle M5 Pro, full request including tokenization:
- INT8: 7.6 ms routing, 8.9 ms extraction, 0.85 GB peak
- FP16: 8.8 / 10.0 ms, 1.58 GB
- FP32: 11.1 / 12.6 ms, 2.55 GB (Python gliner2-mlx: 13.7 / 15.0 ms)

Every precision returns the same labels, spans and offsets as the PyTorch original on 24 reference cases. Most of the speedup came from one change: DeBERTa's relative-position projections depend only on the weights, so they are computed once at load instead of on every request.

I also compared it with TypeSafe's hosted Jev. Worth knowing: the published 60.2% vs 57.6% result is against JevK5, an open reproduction, not Jev itself.

📝 Write-up with animated walkthroughs: https://soniqo.audio/blog/gliner-decide-vs-jev
💻 Code: https://github.com/soniqo/speech-swift
aufklarer 
published a model about 2 months ago