News
Updates, product notes and practical coverage for AI platforms and operators.
AWS has upgraded Amazon Polly’s Generative engine with 14 voices across nine locales, including generally available Hindi and Finnish voices. Existing generative voice names and API calls remain unchanged, so applications receive the new model without code changes. Availability spans 11 commercial and GovCloud regions, including Sydney, while standard and neural engines remain unaffected.
AWS has published a SageMaker AI pattern that streams each completed agent sentence to the Kokoro-82M text-to-speech model instead of waiting for a full answer. Tests reduced transport time to first audio by 52.4 per cent and delivered 2.4-to-3.5-times faster end-to-end starts, while documenting binary transport, buffering and dependency-security trade-offs.
Mistral has released Voxtral 3 speech models and a LiveKit-based voice-agent SDK for real-time multilingual conversations. The family covers transcription, diarisation, end-of-turn detection, expressive speech and tool calling, with EU, US and global APIs plus private deployment options. Pricing starts at US$0.0035 per minute for batch transcription.
Deepgram has ported offline Nova-3 speech-to-text inference to AWS Trainium2, running eight replicas on one trn2.3xlarge instance. The company reports 11–70 per cent more audio processed per dollar for its workload and a further 7.5 per cent throughput gain from a Rust runtime. The results depend on static graphs, custom kernels and workload-specific tuning.
SpaceXAI has released Grok Voice Transcribe 2.0, claiming twice the accuracy of its first transcription model at the same price. The 18 September launch focuses on noisy, multilingual real-world speech and adds diarisation, timestamps and vocabulary prompting. Buyers should test their own accents, environments and error costs rather than treating vendor evaluations as universal performance.
ElevenLabs has launched Eleven v4 and a lower-latency Turbo variant for text-to-speech. The company emphasises emotional range, speaker consistency, multilingual output and professional voice-clone support. Both are available through its creative, agent and API products. Published preference and latency claims need independent testing with the voices, scripts and network conditions a team actually uses.
Google has introduced Gemini 3.8 Live and Live Extended Thinking for real-time voice applications. The models support visual context, multilingual conversation and tool calls that can run while dialogue continues. The Live API release is distinct from rollout across consumer products; benchmark claims and regional access should be assessed against the actual use case.
OpenAI has introduced GPT-Live-1 in the API for more natural real-time voice experiences. The release gives developers a new option for spoken, conversational applications, supporting teams that need responsive audio interaction alongside text-based workflows, with product design, testing, safety controls and clear user correction paths remaining essential.
ByteDance Seed has released SeedRealtime, a full-duplex model that continuously processes audio, video and text while deciding when to speak or act. The model is designed for fluid, interruption-aware interaction, with demonstrations spanning noisy group conversations, visual guidance and proactive assistance. Access and pricing details remain limited.
OpenAI has explained the low-latency architecture behind GPT-Live, the full-duplex voice system now powering newer ChatGPT Voice capabilities. Its design separates continuous audio from slower reasoning and tool work, while an upcoming API is intended to extend the same foundation to developers.
Chatbase now lets organisations connect existing phone numbers to its AI voice agents through SIP trunks. The feature is designed for businesses that want to retain current numbers and telephony infrastructure while routing incoming calls to an agent configured with their own voice, knowledge and behaviour.