September 2026 Summaries

Voice Design enables applications to generate natural-sounding custom voices from text prompts, allowing voice characteristics to be selected dynamically from conversational context rather than limited to preset catalogs or recorded samples. The described customer-support pipeline uses the Jev decision model to classify incoming messages by signals such as tone and urgency, combines those with fixed product-defined persona attributes such as language, age range, accent, and register, and converts them into controlled voice descriptions for Gradium Voice Design. To avoid conversational delays, the system immediately selects an exact or nearest approved voice from a library while background workers generate, audition, and convert new voices for previously unseen specifications. The approach emphasizes small closed label sets, confidence-based defaults, safeguards that prevent user text from directly influencing voice-generation prompts, and deliberate voice changes that preserve consistency within a session. It also recommends keeping agent persona independent of inferred user traits, honoring explicit user voice preferences, and reviewing generated voices for sensitive use cases such as support, while enabling scalable use cases including adaptive narration, game characters, and context-aware assistants.

Coval and Gradium describe a 2026 update to the Coval TTS benchmark that measures perceived time to first audio (TTFA) by combining audio-stream arrival time, including network latency, with any leading inaudible audio before speech begins. The methodological change raised Gradium’s reported TTFA from 171.9 ms to 429.6 ms despite unchanged model and infrastructure performance, because its streams contained a median 225 ms of initial silence. Rather than relying on voice activity detection, which may overlook audible breaths, Coval detects the first audible sound using a normalized RMS loudness threshold over 10 ms windows. The revised metric revealed that models can achieve fast response-arrival times while still feeling slow to listeners if they send silence first, with Cartesia Sonic 3.5 cited as an example. In response, Gradium trained and deployed a new default model that produces audible audio immediately, reducing its median perceived TTFA to 214 ms. Coval notes that results recorded before and after the methodology change are not directly comparable, while providing open-source benchmark code, datasets, methodology, and commit-specific reproducibility information.

Gradium Text-to-Speech is available through LiveKit Inference free until October 9, offering conversational voice capabilities such as voice cloning, sub-250 ms time-to-first-audio latency, and pronunciation support for complex inputs. LiveKit provides an open-source framework and developer platform for building real-time voice, video, and physical AI agents, while its managed Inference service lets developers access models without separate provider accounts, credentials, billing arrangements, or rate-limit management. Gradium can be selected through the `gradium/default` model ID for English, French, German, Spanish, and Portuguese voices with regional accents, and is supported in Python, Node.js, and LiveKit’s browser-based Agent Builder. Built-in text normalization improves spoken handling of phone numbers, dates, codes, and URLs, while teams seeking direct control over billing and limits can continue using the standalone Gradium plugin.

Gradium’s Voice Design feature enables developers to create permanent custom text-to-speech voices from one- or two-sentence descriptions rather than sourcing, consenting, and licensing individual voice actors for cloning. Available through its API and Studio, the tool can generate up to five candidates within seconds and supports requested characteristics such as regional accent, age, gender, pitch, pace, energy, vocal register, and intended application, including customer support, games, audiobooks, and assistants. Gradium reports that in 7,627 blinded pairwise tests conducted with native listeners across English, French, German, Spanish, and Portuguese accent prompts, its voices were preferred in 72.6% of comparisons against ElevenLabs, Inworld, MiniMax, and Fish Audio, with first-place results in every tested language. Its strongest reported accent results included Québécois French, Rioplatense Spanish, Bavarian German, Colombian Spanish, and African Portuguese. A separate Gemini 3.1 Pro evaluation using 50 accent prompts produced the same vendor ranking, though it used independent 1-to-5 ratings rather than human pairwise preferences.

Gradium offers enterprise controls for data residency and retention, addressing where AI inference is processed and whether request content is stored. By default, API requests are routed to the nearest server, while enrolled paid-plan organizations can enforce EU or US session processing through regional endpoints, with requests rejected if they cannot remain in the selected region; responses report the residency location for verification. Organizations may use explicit residency, which applies only to calls made through a chosen regional endpoint, or implicit residency, which pins all traffic to one region and flags incorrectly configured requests. Although session data can be processed in the US under US residency, account information, custom voices, and pronunciation dictionaries remain in Gradium’s EU control plane. Zero Data Retention, configurable through Gradium Studio for paid customers, prevents text, audio, transcripts, and generated outputs from being written to disk after a session, though non-content billing and usage records remain and custom voices are retained until deleted. For organizations requiring data never leave their own environment, Gradium also provides self-hosted AWS SageMaker inference within a customer’s VPC and offers on-premises deployments.