Gradium

What Is Gradium?

Latency is the deciding factor in voice AI — not voice quality, not pricing, not the size of the voice library. When a voice agent pauses for 800ms before it responds, users disengage. Gradium was built around this single insight. Founded in 2025 as a spinout from the Kyutai research lab — the same Paris institution backed by French billionaire Xavier Niel — Gradium came out of stealth in December 2025 with a $70M seed round and one central claim: sub-300ms time-to-first-audio for streaming TTS.

The founding team is credible. Neil Zeghidour (CEO) spent years researching voice models at Google DeepMind and Meta. Olivier Teboul (CTO) was at Google Brain. Alexandre Défossez (CSO) built audio work at Meta. Kyutai itself is where Moshi — one of the first real-time voice-to-voice AI systems — was developed. Gradium is an attempt to take that research and productionise it for developers building conversational AI, voice agents, and real-time translation systems.

As a French company (Gradium SAS), it falls under GDPR by default. The product is genuinely new — as of mid-2026, the company has no track record of enterprise-scale reliability, and some of its positioning reflects early-stage ambition more than proven delivery. That said, the technology is real, the team is serious, and for European developers who need ultra-low-latency voice APIs, there are very few credible EU-headquartered alternatives.

Key Features

Streaming TTS with Sub-300ms Latency

Gradium's headline number is ~220ms time-to-first-audio (TTFA) for streaming text-to-speech. This compares to 400-600ms typical of competitors like ElevenLabs on streaming endpoints. The difference matters practically: at 220ms, a voice agent can respond fast enough to feel like a real conversation. At 600ms, it feels like a telephone call with satellite delay. Gradium achieves this through a combination of a compact audio language model architecture and WebSocket-native streaming — the audio starts playing before the full sentence is synthesised.

The TTS supports English, French, Spanish, Portuguese, and German at launch, with additional languages in development. Multilingual support aligns well with European market needs; French and German neural TTS is a meaningful differentiator for EU-focused teams.

Speech-to-Text API

Beyond TTS, Gradium offers a real-time STT API for transcription. The credit system treats 1 second of STT as 3 credits (versus 1 credit per character for TTS), giving a sense of relative compute cost. STT is available on all plans, including the free tier. Benchmark comparisons published by Gradium itself show competitive accuracy and latency on English speech.

Voice Cloning

Instant Voice Clone is available from the free tier (5 clones included). Pro Voice Clone — a higher-fidelity option requiring a longer sample — unlocks at the M Plan ($340/month). Voice cloning using biometric voice data carries specific GDPR obligations: you need explicit user consent before cloning anyone's voice. Gradium's terms acknowledge this, though the practical compliance tooling is still developer-managed.

On-Device Phonon SDK

The Phonon SDK runs text-to-speech entirely on-device — on a standard smartphone CPU, with no internet connection. No audio is sent to a server. For teams building healthcare apps, offline field tools, or security-sensitive products, this eliminates the data transfer concern entirely. Phonon is also available for embedded deployments. It is a rare capability among voice AI providers in any tier.

Live Speech-to-Speech Translation

Gradium offers live translation between supported languages in real time. This positions it for use cases like international customer support bots, multilingual meeting assistants, or live event interpretation. The feature is relatively new and the breadth of language pairs is still limited, but the underlying architecture — combining low-latency STT, translation, and TTS in a single pipeline — is architecturally sound.

Pricing

Gradium's credit-based system is straightforward: 1 character of TTS equals 1 credit, and 1 second of STT equals 3 credits. The free tier provides 45,000 credits per month — roughly 1 hour of TTS output or 4 hours of STT transcription — with no credit card required. This is genuinely useful for evaluation and small-scale prototyping.

Paid plans start at $43/month for the S Plan (~83 hours of STT), step up to $340/month for the M Plan (~833 hours, plus 5 Pro Voice Clones), and reach $1,615/month for the L Plan (~4,167 hours, 20 Pro Voice Clones). Pay-as-you-go overflow is available at $3.80 per 100,000 additional credits.

There are two structural notes worth flagging. First, all pricing is in USD, not EUR. For EU teams with euro-denominated budgets, this creates FX exposure that accumulates at scale. Second, the pricing is competitive by current market standards — Gradium's per-hour STT rates undercut several established competitors — though as a new company, pricing could shift as they establish market position. Enterprise and on-premise pricing is custom and requires contacting the team.

The Startup Program offers $2,000 in free credits for qualifying seed-funded teams, which is a smart way to build adoption among the developer community that builds the conversational AI products Gradium ultimately wants to power.

EU Compliance & Privacy

Gradium SAS is a French company, which means GDPR compliance is a baseline legal obligation, not a marketing claim. That structural advantage is real. However, the honest picture is more nuanced: the standard cloud API routes through servers in both Europe and the US, meaning confirmed EU-only data residency is not guaranteed on the S, M, or L cloud plans.

For teams that need EU-only data residency, Gradium offers private cloud and on-premise deployment — but these require engaging the enterprise sales process. The on-device Phonon SDK, by contrast, is inherently data-sovereign: audio never leaves the device.

Voice cloning data qualifies as biometric data under GDPR. Developers using Gradium's voice cloning features are responsible for obtaining appropriate user consent before recording and processing voice samples. Gradium's own documentation flags this, though it does not provide consent management tooling — that remains the developer's responsibility.

Being founded in 2025, Gradium has not yet pursued SOC 2 or ISO 27001 certification. For enterprise procurement teams with compliance checklists, this is a gap that will need addressing as the company matures.

Who It's Best For

Developers building real-time voice agents where sub-500ms response time is non-negotiable. Customer service bots, AI call-centre applications, and voice-first interfaces all benefit directly from Gradium's latency advantage.

EU teams needing a GDPR-native voice API who want a French company as the data processor rather than a US-headquartered provider. The legal basis is cleaner, even if EU-only data residency requires the enterprise tier.

Mobile and embedded product teams who need on-device TTS without a server round-trip. The Phonon SDK is purpose-built for this constraint and has few direct competitors among European providers.

If you need a large pre-built voice library, proven uptime SLAs, deep third-party integrations, or a company with years of track record, look elsewhere. ElevenLabs, despite being US-based, has all of these. Gradium trades track record for technical ambition — a reasonable trade for early-adopter teams, not for risk-averse enterprise procurement.

The Verdict

Gradium is a technically credible attempt to build a European voice AI infrastructure company around the one dimension that competitors have not fully solved: latency. The Kyutai research pedigree is genuine. The $70M seed gives it runway to execute. The Phonon on-device SDK is a distinctive capability. The honest caveats are real too: it came out of stealth in late 2025, pricing is in USD, EU-only data residency is not standard on cloud plans, and the integration ecosystem is minimal. For developer teams willing to work with an emerging platform, the technical case is strong. For teams that need certifications, SLAs, and ecosystem depth today, give it another 12 months.

Frequently Asked Questions

Is Gradium GDPR compliant?

Gradium is a French company (Gradium SAS) subject to GDPR by default. The standard cloud API uses servers in both Europe and the US, so EU-only data residency is not guaranteed on standard plans. For strict data sovereignty, Gradium offers private cloud and on-premise deployment. The on-device Phonon SDK keeps all audio on-device with no server contact.

Does Gradium offer a free tier?

Yes. The free tier includes 45,000 credits per month — approximately 1 hour of TTS or 4 hours of STT — with 5 Instant Voice Clones and full API access. No credit card is required. The Startup Program offers an additional $2,000 in free credits for qualifying seed-funded companies.

What is Gradium's time-to-first-audio latency?

Gradium targets sub-300ms time-to-first-audio (TTFA) for streaming TTS, with a published benchmark of approximately 220ms. This is designed for real-time voice agent applications where latency directly determines whether a conversation feels natural or stilted.

How does Gradium compare to ElevenLabs?

Gradium's primary differentiator is streaming latency — it benchmarks faster than ElevenLabs for real-time applications. As a French company, GDPR compliance is structural. ElevenLabs has a larger voice library, broader integrations, more mature enterprise tooling, and a longer track record. Gradium is newer, priced in USD, and still building its ecosystem.

Can Gradium run on-device without an internet connection?

Yes. The Gradium Phonon SDK runs TTS on-device on iOS and Android without any server dependency. It operates on standard smartphone CPUs and is suitable for offline products, edge deployments, and air-gapped environments where sending audio to the cloud is not viable.