Gemini 3.8 TTS voice consistency issue

Hello,

I’m having a problem with gemini 3.8 TTS generation. I noticed it drifts much less than previous models but it also randomly switches voice tone.
I tested without any instruction and a google default voice, created multiple 2 to 7 min audio and in almost all I can hear at least one abrupt change in tone, like if it’s a different voice.

This always happens after a phrase change, so I tried on AI studio to replace periods with commas in order to avoid the issue, but then it drifts a lot and “crashes” at some point, meaning when voice drifts too much it refreshes, probably a safeguard (though I tried this only once).

Then, I tried with multi-speaker mode, one narrator, splitting in groups of 1, 2, 4, 16 phrases. Each still causes a random switch in tone.

Do you have any idea on how to fix the issue? Unfortunately this makes it unusable for my use case

Adding to the above, I also just noticed a voice switch happening within a phrase, and the model adding a word not existing in the text (the second “molecule”, from the example) right at the switch point.

Hello @Thomas_Rossi_Mel ,

Thanks for the detailed report and for sharing the audio sample. We have shared this with the engineering team and are looking into the voice consistency shift on longer audio generations. Could you also share which voice name you selected and the full text prompt used for that sample (and your Project ID via direct message) so we can attach them to the investigation?

Sure,

Voice: Arlo, but same issue with other voices

Model: gemini-3.8-flash-tts, but same issue with Flash-Lite

Not sure how to send the prompt and project via DM, I don’t see the button, do you have an email?

Hi, we’re seeing a similar issue with google/gemini-3.8-flash-tts through OpenRouter.

Our setup:

  • Default voice: Alnilam
  • One request containing a 715-word Bulgarian script
  • One continuous response, approximately 5 minutes 16 seconds
  • No style override or speed/pitch adjustments
  • Input includes short pause tags and Bulgarian syllable-stress marks

The voice sounds noticeably different between passages around 01:04 and 01:12 in the original recording—almost like another person speaking. Both passages come from the same uninterrupted API response, so this isn’t a mismatch between separately generated clips.

We switched from separate scene recordings to one full-script request specifically to improve consistency, but the issue persists.

Is there a recommended workaround for maintaining the same prebuilt voice throughout a longer single-speaker recording?

Hello @Thomas_Rossi_Mel , @Kristiyan_Katsarov ,

DM’ed you for more details.

I got the same problem, one single voice in 10 mins drift other voice around 5 times, it is so annoyed that I can not use gemini flash 3.8 tts for my project like storytelling with one signature voice. I really want to know when the consistency issue was fixed, how i can receive any update about this update or other alternative method for my case, 1 voice with high volume, vibrant and consistency, pls let me know asap thank you in advance!

Hello @sangnguyen0402 ,

DM’ed you for more details.

We see the same voice-consistency problem on gemini-3.8-live (Live API, prebuilt voice Kore). Within a single session, an occasional reply comes out in a clearly different, lower male-sounding voice, then the next reply returns to Kore. In our controlled tests it is rare without video (about 1 in 1,700 replies), roughly 13× more frequent with camera frames streamed into the session, and it never happened on gemini-2.5-flash-native-audio. No reconnects or context compression were involved. We have a minimal reproduction with the bare SDK and can share clips. Is the voice-consistency fix you’re working on for 3.8-flash-tts expected to cover the Live model too?

Following up on our post above: the full write-up for gemini-3.8-live is now in its own topic, with measurements, a minimal bare-SDK reproduction script and pitch plots:

In short, it is the same kind of voice-consistency shift: with prebuilt voice Kore, whole replies come out in one consistent male-sounding voice, in about 2-4% of replies when video frames are streamed into the session and a child or a woman is speaking (about 0.1% audio-only), and in none (0 of 526) on gemini-2.5-flash-native-audio-latest with the same input. We can share WAV samples by direct message.

Hello @sangnguyen0402 ,

Thank you for sharing your request payload and script. We are looking into this issue and will share an update once we have more information.

the same issue. it’s so annoying. hope it will be fixed soon

I would consider Gemini 3.8 Flash TTS as one of the biggest disappointments in AI. For the cost, its the best sounding and performing. But it parks in a handicapped zone. The drifting. It even interprets texts, add words, hallucinates. This isn’t your typical TTS model. I dare to even call it that. Switching to Elevenlabs is a no brainer. Even the el cheapo Kokoro is at least far more reliable than Gemini. You would think Google would be on top of this but it reminds me of Microsoft. They were notorious for leaving things broken while pretending to fix it. As my dad always said, you can’t fix stupid.

I’m experiencing severe voice drift and voice inconsistency with Gemini 3.8 TTS, even when using Google’s official preset voices.

Even when I use the exact same preset voice, identical API parameters, and the exact same text, repeated generations can produce completely different-sounding voices. The inconsistency is unpredictable and extremely frustrating.

The problem becomes even worse with long-form text. Voice drift happens frequently during a single generation, making it nearly impossible to maintain a consistent voice throughout a long narration. Sometimes, a female voice suddenly turns into a male voice in the middle of the audio!

What’s particularly frustrating is that Gemini 3.8 TTS is actually excellent in terms of naturalness, audio quality, and pricing. It has enormous potential. However, this bizarre and severe voice consistency issue makes it practically unusable for real-world production.

Please prioritize fixing this bug as soon as possible!

Frankly, I find it hard to believe that proper quality testing was conducted before this model was released. This is such a fundamental and easily noticeable problem that I’m genuinely shocked it wasn’t caught during internal testing.

How could such an obvious and critical issue make it into a public release?

I really hope the engineering team takes this seriously and provides a fix soon. Voice consistency is not an optional feature for a TTS model — it is one of the most basic requirements.

Same here, for days. It drives me crazy because the voice is very good, but I have to repeat generation even for shorter takes 10 times or more to find a non drifting version. It’s not reliable and I really hope Google will fix this and not just pretending to do so…

Hello Everyone,

Thank you for the additional reports. We are looking into this issue and will share an update here once it is resolved.

Addition: we now also see frequent audio dropouts (gemini-3.8-flash-tts, voice Achird, German, PCM, no style prompt). Mid-word the voice breaks up for 10 to 20 ms, like a small crackle. It is in the raw output, via AI Studio, Vertex and OpenRouter alike, in more than half of our takes, regardless of request length. It’s a bit embarrassing how much budget we literally burned for useless audio takes… when we started using that voice, weeks ago, none of the problems occured. Actually it’s a shame, to pay money for that quality at the moment.

We’re seeing the same problem with Gemini 3.8 Flash TTS, using a Voice Design voice through the Gemini API to narrate long educational audio. The speaker stays the same, but at certain points it sounds like a different recording session; the loudness steps up or down and the tone gets brighter or darker. The direction varies from take to take. One take got louder and brighter at a given point, and another got quieter and darker at the same point.

We think we can show where it happens. It lines up with the internal segments reported in the usage data that comes back with each response (model_invocation_token_counts):

  • A 2,019-word request ran as 7 invocations, roughly 290 words each.
  • The first 301 words of the same text, sent on their own, ran as 1 invocation (392 text tokens).
  • The remaining 1,718 words, sent on their own, ran as 6.
  • In the 2,019-word request, the first invocation’s text prompt was 387 tokens, so the first segment is those same 301 words.
  • Every invocation’s prompt also includes 840 audio tokens, which we take to be the voice reference. The later prompts are larger (about 820 to 1,040 text tokens), which suggests each one is also given some of the preceding text.

The change we hear falls exactly on that first boundary, at the start of the paragraph after word 301. It does so in the 2,019-word request and in a separate 605-word request that begins with the same text. How noticeable it is varies between takes: across 10 takes we measured at that point, the level stepped up by 0.8 to 2.1 dB in 7. But when it is audible, that is where it is. The same passage sent on its own, in 12 short requests with four different wordings, never showed it, so the text itself isn’t the cause.

Our reading is that each internal segment starts again from the voice reference with only the preceding text, not the preceding audio, so its recording character doesn’t always match the segment before. We tried to continue from earlier audio with previous_interaction_id, but that is refused (“Audio input modality is not enabled for this model”). Splitting the text into shorter requests ourselves didn’t help either, because the joins between requests sounded the same way.

Happy to share the project ID and request IDs by DM if that would help the engineering team.

any news in here? this model is useless for long audio text to speech. I insist keeping using it as the voices are so great but how??

I get the suspicion, that nothing is going to happen here… I do not really see any support from Google here, just some “we look into this” posts, but nothing else. We already switched with a new project to Elevenlabs. The voice is not as good, but at least it’s consistent and we save 90% of time that we wasted just for quality control and endless regenerations (which were, of course, not for free!) with the Google modell. Actually Google should pay back all the wasted budget for broken outputs. But of course we all know, they will not do that (nor fix the problems)