AWS adds 14 Polly voices and PyTorch support for Trainium chips

The week in AWS

Amazon Web Services' weekly roundup for 26 October 2026 lists several launches aimed at voice, AI model training and security. The post is written by AWS's Daniel Abib ^[1].

Better voices from Amazon Polly

The Amazon Polly Generative engine has been upgraded with a larger, more expressive catalogue: 57 generative voices across 11 languages and 25 locales. There are 14 new voices across nine locales, and Hindi and Finnish are generally available for the first time. Two of the new voices are Brazilian Portuguese, named Leo and Carolina ^[1].

AWS calls it a drop-in upgrade: voice names, the same SynthesizeSpeech call and the price all stay the same, while the output becomes more natural ^[1]. For businesses, that means existing apps can sound better without code changes.

PyTorch comes to Trainium

AWS announced TorchNeuron, a public-beta, open source PyTorch backend for its Trainium chips. It lets developers run a model on Trainium by switching to .to('neuron'), which AWS describes as a configuration change rather than a rewrite, with eager execution and torch.compile intact. It works on Trainium2 and Trainium3 instances and lands alongside AWS Neuron 2.33 ^[1]. The aim is to make it easier to try AWS's own AI chips without rebuilding software.

Security: AWS Continuum

The biggest launch, according to the post, is AWS Continuum for code vulnerabilities, now generally available. It builds a context graph from a customer's repositories and AWS environment, assigns each finding an evidence level, and reproduces exploitable findings in an isolated sandbox with proof-of-concept code. It ranks issues by how close the attack path runs to business-critical systems, then opens the fix as a pull request. Every remediation currently requires human approval ^[1]. It is available in US East (N. Virginia) and US West (Oregon).

AWS says its honeypots saw a PHP2Shell exploitation attempt on 17 October, four days before two related vulnerabilities were publicly disclosed ^[1]. AWS draws the conclusion that a patch window measured in days assumes a delay that no longer exists.

Other launches

• Amazon SageMaker AI training sessions are generally available, giving control of the post-training loop, including custom loss functions and reward signals, while AWS handles GPU allocation and recovery. Customers pay for tokens processed ^[1].
• Structured outputs for Anthropic's Claude models on Amazon Bedrock return schema-compliant JSON without custom parsing or retry logic, which suits tasks such as extracting fields from financial reports or routing support tickets ^[1].
• Amazon ElastiCache Serverless for Valkey now supports online migration with zero downtime ^[1].
• Amazon ECS Gateway lets platform teams define shared network settings while application teams opt in. It is available at no additional charge ^[1].

What it means

Taken together, the launches show cloud providers making AI cheaper to run, easier to adopt and safer to operate. Teams that build voice features, train custom models or manage large code bases may find something useful here. As with any vendor roundup, the descriptions come from AWS itself, so check regional availability and pricing before planning around a feature.

AWS re:Invent runs from 30 November to 4 December in Las Vegas ^[1].

Source: AWS — AWS Weekly Roundup: Amazon Polly Generative engine, TorchNeuron for Trainium, and more (October 26, 2026)