Google recently introduced its Gemini 2.5 text-to-speech (TTS) model, drawing attention across the voice AI community. But how does it actually perform when measured against established models like ElevenLabs’ Multilingual V2?

At Podonos, we believe performance claims should be backed by transparent, data-driven analysis. That’s why we conducted a head-to-head evaluation of Gemini 2.5 Flash and ElevenLabs’ latest multilingual model.

Key Findings

1. Overall Performance

Both models scored similarly in user preferences, but ElevenLabs edged ahead slightly in overall quality.

2. Weakness in Address and Number Pronunciation

Both models showed notable difficulty handling addresses and numbers—highlighting a common challenge in TTS robustness.

3. Dialog and Named Entity Handling

Gemini underperformed in dialog-based speech, especially when pronouncing celebrity names and medical terms, suggesting gaps in real-world context handling.

4. Diversity and Inclusion

Gemini showed a notable imbalance in voice quality across genders, performing significantly better on male voices than female voices. This raises concerns around bias and inclusivity in synthesized speech.

You can find more insights in the full reports below.

📝 Naturalness comparison
📝 Preferences

Why This Matters

As voice AI becomes a core interface in digital experiences, accurate and fair performance evaluation is no longer optional. Models must be tested not only for naturalness and clarity, but also for consistency across diverse content and speaker profiles.

At Podonos, our goal is to make this kind of rigorous evaluation accessible to any AI team. Whether you're launching a new model or refining an existing one, Podonos helps you identify blind spots, benchmark against competitors, and make confident improvements.