Gradium

What Gradium is

Gradium is a voice AI platform aimed at developers building real-time agents rather than at people making one-off voiceovers. It ships four model families behind one API: text-to-speech, speech-to-text, voice cloning, and live speech-to-speech translation. The pitch is latency and deployment flexibility, not catalog size.

Two things separate it from the crowded TTS market. First, it publishes a first-audio figure of about 200 ms over a bidirectional WebSocket API that accepts input before the ready signal returns, with telephony audio formats supported directly, which is the specific combination phone agents need. Second, it offers a 100M-parameter on-device TTS model that runs fully offline on CPU, including Apple, Android and Raspberry Pi, covering every supported language and any voice. Very few voice vendors ship a serious offline path.

►

Pricing

Credit-based, with six tiers. Verified from Gradium's pricing page in August 2026:

• Free, $0: 45,000 credits/month, Studio access, 5 instant voice clones. No commercial use.
• XS, $13/month: 225,000 credits, API access, commercial use enabled
• S, $43/month: 900,000 credits, higher concurrency limits
• M, $340/month: 9M credits, 5 pro voice clones
• L, $1,615/month: 45M credits, 20 pro voice clones
• Enterprise, custom: unlimited credits, custom concurrency, full feature set

Extra credits are sold in 100,000 blocks and get cheaper as you scale: $6.90 on XS, $5.00 on S, $4.00 on M, $3.80 on L. Gradium also advertises $2,000+ in free credits for startups through a grant program.

The credit math is the part worth reading twice. Text-to-speech bills 1 credit per character, and Gradium puts roughly 45,000 characters at about an hour of audio, so the free tier is about one hour of speech per month. Speech-to-text runs 3 credits per second (about 10,800 an hour), speech-to-text translation 4 per second, and speech-to-speech translation 30 per second. That last number is an order of magnitude above transcription, so a live-translation workload burns a tier far faster than the headline credit counts suggest. Model your actual mix before choosing.

Where Gradium is the strongest pick

Real-time voice agents where latency is the product, and any deployment where audio cannot leave the device or the building. The offline CPU model plus self-hosted and on-premise options make it a credible option for regulated or air-gapped work, which rules out most of its API-only competitors immediately.

The integration story supports that: official Python and Rust SDKs, and first-class support for LiveKit and Pipecat, the two frameworks most real-time voice agents are actually built on, plus an open-source bot framework of their own.

Best for

Developers building latency-sensitive voice agents, telephony products, and teams with on-premise or offline requirements. Not aimed at marketers wanting a quick voiceover, and not a fit if you need broad language coverage.

Key features

• Text-to-speech, speech-to-text, voice cloning and live speech-to-speech translation in one API
• Bidirectional WebSocket streaming with a stated first-audio time of about 200 ms
• Instant voice cloning from 10 seconds of audio, pro clones for higher fidelity, or voice design from a prompt
• On-device TTS: 100M parameters, fully offline on CPU (Apple, Android, Raspberry Pi)
• Deployment as cloud API, dedicated instances, self-hosted or on-premise
• Official Python and Rust SDKs, first-class LiveKit and Pipecat integrations
• Telephony audio formats supported natively

Best-fit use cases

• Phone-based AI agents where round-trip latency decides whether the conversation feels natural
• Live translation in conversation, budgeted carefully against the 30-credits-per-second rate
• Offline or on-device speech in apps, kiosks and hardware with no reliable connectivity
• Regulated deployments where audio must stay on-premise

Pros

• Genuinely low stated latency with an API designed for streaming, not request-response
• Offline on-device model is rare in this category and unlocks air-gapped use
• Deployment flexibility from cloud through to on-prem
• LiveKit and Pipecat support means less glue code for real agent stacks
• Cheap entry point at $13/month once you need commercial rights

Cons

• Five languages only (English, French, Spanish, Portuguese, German), narrow next to rivals advertising 30+
• Free tier bars commercial use, so it is an evaluation tier rather than a real free plan
• Credit pricing makes cost forecasting harder than per-character or per-minute billing
• Live translation at 30 credits/second can consume a tier far faster than the headline numbers imply
• The ~200 ms figure is the vendor's own benchmark, not an independent measurement
• Big jump from S ($43) to M ($340) with little between for mid-sized workloads

What to watch out for

Check the language list before anything else. Five languages is the single most likely reason Gradium will not fit, and no amount of latency advantage compensates if your users speak something it does not support.

Do the credit arithmetic against your real workload. A product that is mostly TTS behaves very differently on these tiers than one doing live translation, where the per-second rate is ten times higher. Work out monthly seconds by mode, not total credits.

Treat the latency claim as a starting point. Vendor-published figures are measured under favorable conditions. If latency is why you are choosing Gradium, benchmark it from your own region and network before committing.

Is Gradium worth it?

For its target use case, it is a serious option: the combination of streaming-first design, an offline CPU model and on-premise deployment is not something most voice APIs can match, and $13/month for commercial rights is a low bar to test it properly. The honest limiter is language coverage. If English, French, Spanish, Portuguese or German cover your users, Gradium deserves a place on the shortlist next to Cartesia and ElevenLabs. If they do not, the decision is already made for you.

Worth comparing against our roundups of AI text-to-speech tools and AI voice agents before you commit.

FAQ

Pricing, credit rates, latency and language claims verified August 2026 from gradium.ai and gradium.ai/pricing. Gradium has no affiliate program and no paid placement with ToolChase; the links above are plain, untracked links. Gradium is listed under our free editorial badge program, which does not affect this review or any ranking.