Kimi K3.1 Pro on AMD Instinct MI355X offers Day 0 support with 2.8T parameters and 1M token context. It runs on 8 × MI355X using ATOM and vLLM.
Key points
Params: 2.8T
Context length: 1M tokens
Deployment: 8 × MI355X with ATOM and vLLM
Sources
- Original post
- amd.com - Linked in the post
GPT-Live 1 (Terra, low) ranks high overall despite low audio score
GPT-Live 1 (Terra, low) achieved the second-highest average overall rating in the DIAL benchmark while placing near the bottom for audio quality. Across about 8,000 consumer and business calls, helpfulness ratings aligned more with overall scores than with naturalness or audio quality.
Key points
GPT-Live 1 (Terra, low) is second-highest overall rating in DIAL benchmark, third-to-last audio quality.
All tested configurations averaged at least 3.60/5 on audio quality.
Sources
ElevenLabs uses Silencio for consent-cleared voice data
ElevenLabs now uses Silencio for consent-cleared voice data, marking a partnership with the leading voice AI company. The post calls the data sourcing engine built by Silencio over the years an incredible validation of its quality.
Key points
ElevenLabs now uses Silencio for consent-cleared voice data.
Silencio has built a data sourcing engine over the years and produces high-quality data.
Sources
Soniox STT v6 launches with agent voice echo cancellation
Soniox STT v6 is the world’s first speech-to-text model with agent voice echo cancellation, keeping listening while the agent speaks. It filters the agent’s playback from recognition and works across 60+ languages.
Key points
Agent voice echo cancellation automatically filters the agent’s playback out of recognition.
Model continues to listen during agent speech and recognizes user interruptions.
Agent text input streams the agent’s outgoing text to provide context for both sides of the conversation.
Sources
- Original post
- Soniox Introducing Soniox STT v6 - Linked page
Audio-to-MIDI plugin – transcribe samples to clean MIDI
Audio-to-MIDI plugin lets you drop a sample, keep suggested instruments, hit Transcribe and get clean MIDI parts. It offers a three-step workflow to avoid DAW conversion.
Key points
Drop the sample onto the Audio-to-MIDI plugin
Keep the suggested instruments and hit Transcribe
Drag each part into your project
Sources
SurrealDB notes app on Phonon built for on-device analysis
SurrealDB built a notes app on Phonon that analyzes your notes and reads the findings back to you, all on-device. Phonon combines on-device TTS with a SurrealDB Agent Memory. The project is open source.
Key points
Notes app built on Phonon
Phonon provides on-device TTS and SurrealDB Agent Memory
Sources
- Original post
- github.com - Linked in the post
SoundBoost Music Video Generator adds auto retakes for lip sync
SoundBoost Music Video Generator now includes auto retakes that automatically switch the underlying lip sync model until the video matches, delivering properly lip-synced output. The feature is offered at no extra cost.
Key points
Auto retakes fix lip sync inconsistencies in video models.
The system silently switches the lip sync model until everything matches.
Sources
Verda Inference preview offers day-0 Kimi K3.1 Pro access
Verda Inference preview offers day-0 Kimi K3.1 Pro access via a single API for open-weight models. The service is in partnership with Kimi Moonshot.
Key points
Preview launch: day-0 access to Kimi K3.1 Pro
Single API serves frontier open-weight models
Sources
- Original post
- verda.com - Linked in the post
