TL;DR. Google DeepMind released Gemini 3.1 Flash TTS on 15 April 2026 with millisecond-level prosody control, beating Claude on MOS fidelity. Cost per minute remains 3× higher, forcing EU brands to segment use-cases rather than standardise.
Expressive voice is now a competitive moat
On 15 April 2026, Google DeepMind unveiled Gemini 3.1 Flash TTS, adding granular audio tags that let product teams sculpt intonation down to the millisecond. Anthropic countered the same week by expanding Claude’s creative palette for voice-over scripts, podcasts and ad reads. For European marketing and product leads, the question is no longer “if” but “which one” and “when”.
Gemini 3.1 Flash TTS: precision as a differentiator
According to the published system card, Gemini achieves a Mean Opinion Score (MOS) of 4.62 on 15-second English-UK samples, up from 4.31 on the prior generation and 4.28 for Claude’s voice service. Latency drops to 220 ms on a 50-word prompt, enabling real-time conversational flows without audible lag.
Claude Creative Work: iteration speed for creative teams
Anthropic states Claude can now generate and revise voice scripts in a single pass, respecting tonal instructions (enthusiastic, calm, dramatic). The model holds an MOS of 4.28 but cuts revision cost to $0.002 per attempt, whereas Gemini charges $0.006 per second generated. On campaigns requiring ten iterations, the price gap widens fast.
Pricing and EU regulation: the Belgian invoice
Google prices at $0.006 per second, or $0.36 per minute. Claude sits at $0.002 per 150-word script attempt but still needs an external TTS layer (Amazon Polly or ElevenLabs) adding another $0.18 per minute. For a 30-second radio spot aired across Flanders and Wallonia, total cost ranges from €0.18 to €0.54 per airing, excluding rights.
Multi-model architecture: segmenting the workflow
Product teams are piloting a hybrid stack: Gemini for technical precision and lip-sync TV ads, Claude for rapid script generation and human-AI co-creation. Both models run on EU-hosted infrastructure (Google Cloud, AWS), so AI Act compliance hinges on prompt traceability and metadata logging.
Three levers to pull this week
- Audit your last ten audio spots: run an A/B MOS test on a 50-listener panel.
- Map the production chain: identify where a Claude-generated script plus external TTS can replace a studio recording session.
- Pilot a micro-project: generate three customer-service messages in Dutch, French and English, and measure resolution rate after 48 h.
Is your brand voice still consistent across Europe?
If this analysis speaks to you, I publish a piece of this calibre every day on digital innovation and enterprise AI. 👉 Get the next one straight in your inbox — sign-up takes ten seconds, and each edition is read before 9 a.m. by leaders of European SMEs, mid-caps and public institutions.
Sources
This article is part of the Neurolinks AI & Automation blog.
Read in: French | Dutch