Gradium

Gradium is a voice AI platform for developers building products that need to speak, listen, or translate during a live interaction. Its API covers text-to-speech, speech-to-text, speech-to-speech translation, voice cloning, and voice design, with a browser based studio for trying the models before wiring them into an application. The focus is not simply producing an audio file after a request. Gradium is designed for conversational systems where the first audible response, stable streaming, and knowing when a person has finished talking all affect whether the interaction feels natural. That makes it relevant to voice agents, phone workflows, accessibility tools, and other products that work with speech as it happens.

Text-to-speech can run through a streaming WebSocket so an application can send text as it is generated and receive audio before the full passage is ready. The product pages emphasize fast time to first audio, no leading silence, telephony formats, and steady latency as concurrency rises. Developers can choose from a voice library, create a new voice from a written description, or make an instant clone from a short recording. Voice Design accepts details such as perceived age, accent, texture, and mood, then creates a reusable voice in the studio. Instant cloning works from ten seconds of audio without a separate training step. More involved professional cloning is available on larger plans, while all custom voice use remains subject to moderation and the service terms.

On the listening side, Gradium streams speech-to-text results and includes semantic turn detection in the same model. Instead of treating any silence as the end of a turn, it sends inactivity probabilities as the conversation develops, giving an agent more context for deciding when to respond. Teams can adjust the transcription delay to favor faster output or higher accuracy and can boost a custom vocabulary of product names, people, or specialist terms without retraining a model. The documented behavior keeps words stable after they appear, which is useful for captions and interfaces where rewriting earlier text would cause visible jumping. REST and WebSocket interfaces are documented, along with recipes for browser audio, telephony, turn taking, and LLM output sent into speech.

Live translation uses a direct speech model rather than requiring a separate transcription, text translation, and synthesis chain. The service listens in one supported language and produces speech in another while the speaker is still talking. Gradium currently documents English, French, German, Spanish, and Portuguese across its main speech models, including regional variants and mid-sentence language switching. Any of the five can translate into any of the others, and the output can use a catalog voice or a custom clone. This is narrower language coverage than broad transcription services offer, but the common API across speaking, listening, translation, and custom voices reduces the number of separate vendors a team has to combine for a multilingual conversation.

It suits engineering teams that need a programmable voice layer more than creators looking only for a one-off narration download. A support agent can combine streamed transcription, turn detection, an LLM, and synthesized replies. A calling product can use telephony audio formats, while a translation experience can carry a selected or cloned voice into another language. Gradium documents deployment through its cloud by default, with dedicated, marketplace, self-hosted, and private infrastructure options discussed for larger requirements. Its public materials also state zero data retention and ISO 27001 certification for the managed service. Those deployment and compliance details matter most to teams handling sensitive conversations, though buyers should still review the current trust materials and terms for their own workload.

Access is freemium. The free plan includes 45,000 credits, Studio and API access, and up to five custom voices, but it doesn't include commercial use. Paid tiers begin with larger credit pools, higher concurrency, commercial use, and many more custom voices, while enterprise terms are custom. Credits are consumed differently by text-to-speech characters, transcription time, and translation time, so expected cost depends heavily on the mix of operations rather than one simple hourly rate. Voice Design itself is included across plans, and the studio can be tried without a credit card. The submitted link pointed to a Voice Design demo, but the canonical product is the broader Gradium platform at gradium.ai, where the full API, studio, documentation, and pricing are presented together.