Soniox STT v6 hits 0.71% pooled semantic word error rate and delivers transcripts 266 ms after speech ends. The top three Pareto-frontier models are also Soniox real-time STT models.
Key points
Pooled semantic word error rate: 0.71%
Median transcript latency: 266 ms after the speaker stopped
Sources
- Original post
- soniox.com - Linked in the post
Livekit 89% call accuracy, 73% tasks perfect 5 runs
Livekit achieved 89% correct calls in a test of six voice AI platforms across 90 phone tasks run five times each. It answered all five attempts correctly on 73% of tasks. Across all platforms, the “right every time” metric was 15 to 28 points lower than “right once”.
Key points
@livekit got 89% of calls right.
On 73% of tasks it got the call right all 5 times.
“Right every time” ran 15 to 28 points below “right once” across all six platforms.
Sources
Phonic-1 and Realtime v1 mini rankings released with scores
Phonic-1 and Realtime v1 mini appear on DIAL with overall preference rankings. Phonic-1 is ranked #2 with a 4.04/5 score, while Realtime v1 mini is #12 with a 3.60/5 score. Rankings are based on live multi-turn conversations and include breakdowns by conversation type and quality dimensions.
Key points
Overall preference: Phonic-1 #2, 4.04/5
Overall preference: Realtime v1 mini #12, 3.60/5
Sources
OASYS multimodal simulation adds test scenarios for AI agents
OASYS introduces multimodal simulation to test AI agents against past customer conversations, background noise, bad phone lines and frustrated callers across chat and voice. The feature lets teams release agents with confidence for real-life scenarios.
Key points
Feature: multimodal simulation in OASYS.
Test scope: past conversations, background noise, bad phone lines, frustrated callers, chat and voice.
Sources
Music Video Generator rewrite & reshoot now live, 1-click scene change
Music Video Generator adds a “rewrite & reshoot” feature that is now live. It replaces the multi-step process with a single click to change a scene.
Key points
Feature: rewrite & reshoot is now live in the music video generator.
Interaction: changing a scene now requires only 1 click.
Sources
Soundverse v9 custom mode creates Bhojpuri R&B song
Soundverse v9 in Custom mode was used to generate “Neend Na Aawe,” a Bhojpuri alternative R&B track with 80s synth-pop elements. The song includes gated drums and a falsetto hook and follows its lyrics.
Key points
Model: v9, Custom mode
Style: 80s synth-pop R&B, gated drums and a falsetto hook
Sources
- Original post
- youtu.be - Linked in the post
- soundverse.ai - Linked in the post
