# Gradium opens TTS beta for phone numbers, emails and financial codes > Source: > Published: 2026-07-30 19:20:47+00:00 [Gradium](https://gradium.ai), the voice AI developer led by co-founder and CEO [Neil Zeghidour (@neilzegh)](https://x.com/neilzegh), opened a public beta on July 30th for a text-to-speech model designed to pronounce the strings that routinely break voice agents, including phone numbers, email addresses, IBANs and time expressions. Gradium announced the release [in a post on X](https://x.com/gradiumai/status/2082895563712815458?s=46) and made the model available through its API under the name `gradium-tts-beta` . ([gradium.ai](https://gradium.ai/blog/gradium-tts-public-beta)) [https://x.com/gradiumai/status/2082895563712815458?s=46](https://x.com/gradiumai/status/2082895563712815458?s=46) The beta continues work Gradium began with a production TTS upgrade in June. Gradium says the new model can process difficult inputs directly, without developers maintaining a separate layer of prompts, regular expressions and text-normalization rules. The target cases extend beyond the headline examples to account handles, measurements, reference codes, license plates, URLs and ordinals. The production model remains the default, so customers must explicitly opt into the beta. ([gradium.ai](https://gradium.ai/blog/gradium-tts-public-beta)) That focus reflects Zeghidour's route into the market. Before starting Gradium, he worked on generative audio at Google Brain, Google DeepMind and Meta. He founded Gradium with CTO Olivier Teboul, Chief Coding Officer Laurent Mazare and Chief Science Officer Alexandre Defossez. The group previously worked on Kyutai, the Paris research lab that released real-time conversational speech systems, and Gradium says its founders helped develop and open-source neural audio codecs and audio language models. ([gradium.ai](https://gradium.ai/blog/gradium-100-million-funding-nvidia)) ### Removing a production dependency Text-to-speech systems can sound natural on ordinary prose while failing on the structured information that carries the most operational risk. A misread callback number can end a customer-service interaction. A spoken email address becomes useless if the model mishandles punctuation. Financial identifiers and reference codes must preserve every character and the grouping listeners expect in their language. Developers commonly address those failures before inference, rewriting inputs into phonetic or expanded forms. That adds another component to the voice stack and creates language-specific edge cases that engineering teams must test and maintain. Gradium's stated goal for the beta is simpler: applications should be able to send text directly from a database or language model and receive a usable spoken version without preprocessing. ([gradium.ai](https://gradium.ai/blog/gradium-tts-public-beta)) The model is available through Gradium's [Python SDK](https://docs.gradium.ai) and real-time WebSocket API. Developers select it by passing `gradium-tts-beta` as the model name. Existing Gradium voices, including custom voices, carry over without a migration. Gradium has limited the beta to API users rather than making it the default in production, giving customers a controlled path to test the model against their own traffic. ([gradium.ai](https://gradium.ai/blog/gradium-tts-public-beta)) Gradium is offering 1 million credits to each beta participant who submits feedback based on a production case. At Gradium's [published conversion rate](https://gradium.ai/pricing) of one credit per TTS character and roughly 45,000 characters per hour of generated audio, that incentive represents about 22 hours of synthesis. Participants must provide the input text, language, voice and expected output, giving Gradium a stream of failure cases to use in refining the model. ([gradium.ai](https://gradium.ai/pricing?utm_source=openai)) The feedback offer is central to the release. Structured strings have too many regional formats, abbreviations and pronunciation conventions to cover with a clean benchmark alone. Gradium is using customer workloads to find the long tail, while asking developers to judge the beta on the exact inputs tied to support calls, account verification and other live applications. ### A heavily funded voice stack Gradium launched publicly in December 2025 with a $70 million seed round led by [FirstMark](https://www.firstmark.com/) and [Eurazeo](https://www.eurazeo.com/en?trk=public_post-text). On July 8th, Gradium said it had extended the round to $100 million and added [NVIDIA](https://investor.nvidia.com/home/default.aspx?ncid=ref-inpa-842809) as an investor. Gradium said the financing would fund research, product development, international expansion and a San Francisco Bay Area office. ([gradium.ai](https://gradium.ai/blog/gradium-100-million-funding-nvidia)) The capital gives Gradium room to compete in a voice market that includes ElevenLabs, Cartesia, Inworld and voice products from large model vendors. Gradium has concentrated its pitch on real-time infrastructure: speech generation, transcription, translation, voice cloning and turn detection that can be assembled into interactive agents. Gradium says it began generating revenue within weeks of launch and serves customers in healthcare, customer experience, media, AI agents and consumer applications, though the figures remain company-reported. ([gradium.ai](https://gradium.ai/blog/gradium-100-million-funding-nvidia)) The public beta narrows that broad platform strategy to a specific engineering problem. Gradium is betting that voice providers will compete on whether their models survive ordinary business data, not simply whether a demonstration sounds human. Phone numbers, email addresses and bank identifiers make for less dramatic demos than expressive cloned voices. They decide whether a voice agent can complete the task it was deployed to handle.