Gemini 3.8 text-to-speech
Google’s new Gemini 3.8 text-to-speech model is being received as technically impressive, competitively priced, and in some cases on par with leading commercial systems like ElevenLabs, especially for long-form audio such as audiobooks. Commenters highlight emerging workflows that mix commercial and open-source tools (Qwen, Kokoro, Fish Audio, Higgs, etc.) to get better prosody, local control, and lower costs, while noting that many voices still sound subtly “AI” and often over-emotional. There are also concerns about fragmented availability across Google’s platforms, the ethics and safeguards around voice cloning, and the broader impact on voice actors and game/audio production.
Local and open‑source TTS and audiobook tools
- Several people are building local audiobook / audio‑drama apps (e.g., full‑cast emotive readers with quotation attribution, voice design, and stage‑direction‑style controls).
- Suggested pipelines: use one model for “voice design” and another for final rendering.
- Multiple local tools are recommended: Kokoro (via a Python frontend), Alexandria Audiobook, pdf‑narrator, PocketTTS, audio.cpp, etc.
- Users value “no cloud, no tokens” and running on consumer hardware (M1, mid‑range GPUs).
Model comparisons and perceived quality
- Gemini 3.8 TTS is widely praised as “damn good” and comparable to ElevenLabs v3 in expressiveness.
- Others still prefer open models like Fish Audio, Higgs, Qwen3 TTS, BreezeTTS for better prosody or specific use cases.
- Some say the examples ignore key prompt cues (e.g., laughter tags) and that outputs don’t precisely match directions.
- Complaints that many models overact: exaggerated emotion, random emphasis, “angry” or “insufferable” for long conversations.
- Desire from some users for restrained, consistent “computer voices” with strong pronunciation control.
- British accents are still reported as weak and “fake‑sounding.”
Voice cloning, consent, and abuse concerns
- Voice replication from short samples is seen as already commoditized; “cat is out of the bag.”
- Users report creating highly convincing clones of themselves and celebrities with open tools in hours.
- Concerns about scams, impersonation, and how long consent recordings are stored; skepticism that consent‑verification can’t be defeated by synthetic voices.
Pricing and economics
- Gemini 3.8 TTS pricing (~$0.27–$0.81 per audio hour, cheaper in batch, promo until year‑end) is viewed as very competitive vs ElevenLabs.
- For large books (e.g., 250k words), cost differences determine whether users will adopt the service.
Platform fragmentation and integration gaps
- Frustration that Google models and capabilities differ across consumer, prosumer, Workspace, and GCP; availability and features are inconsistent.
- Users want the new TTS to power products like Play Books, Android/Android Auto, and better Edge/Chrome‑like “read aloud” features; current built‑in voices are considered outdated.
- Rollout feels rough: region restrictions, missing voice types/use cases, need to set up billing before trying, and unclear or scattered documentation.