Gemini 2.5 Pro TTS – Recent audio quality degradation with random micro-cuts / glitches

Google AI Studio
1

I have recently noticed a degradation in the audio quality generated with Gemini 2.5 Pro TTS in Google AI Studio.

The generated speech now occasionally contains very short interruptions, micro-cuts, or discontinuities in the audio. They sound like tiny sections of the waveform are missing or being abruptly interrupted.

This behavior was not noticeable, or was significantly less frequent, in previous generations using the same model and similar prompts.

Model

Gemini 2.5 Pro TTS
Google AI Studio

Problem

During TTS generation, the resulting audio sometimes contains:

  • Very short audio cuts or interruptions.

  • Small discontinuities between phonemes or words.

  • Occasional unnatural transitions in otherwise continuous speech.

  • A noticeable reduction in perceived audio consistency compared with previous generations.

The problem does not necessarily occur at the same position every time, which makes it seem related to the generation process rather than to the input text itself.

Steps to reproduce

  1. Open Google AI Studio.

  2. Select Gemini 2.5 Pro TTS.

  3. Generate speech from a normal paragraph or narration.

  4. Listen carefully to the generated audio.

  5. Repeat the generation several times using the same text.

In some generations, small cuts/glitches can be heard even though the text and settings have not changed.

Expected behavior

The generated speech should be continuous and smooth, without missing audio samples, sudden micro-pauses, clicks, or interruptions between phonemes.

Actual behavior

Some generations contain short but clearly audible discontinuities or micro-cuts that reduce the quality of the generated speech.

This is particularly noticeable when using the audio for narration, where even very small interruptions can be distracting.

Frequency

Intermittent, but frequent enough to be noticeable across multiple generations.

When I started noticing it

Approximately around September 2026.

The issue appears to be more noticeable recently compared with earlier generations made with Gemini 2.5 Pro TTS.

Additional information

It would be useful to confirm whether there have been any recent changes to:

  • The Gemini 2.5 Pro TTS inference pipeline.

  • Audio post-processing.

  • Audio encoding or WAV generation.

  • Streaming/chunk reconstruction.

  • Sampling or generation parameters used internally by Google AI Studio.

I can provide example audio files where the discontinuities are clearly audible if needed.

Thank you.

2

Hello @JL_Garcia ,

We have just launched the new Gemini 3.8 Flash TTS (gemini-3.8-flash-tts) and Gemini 3.8 Flash-Lite TTS (gemini-3.8-flash-lite-tts) models: Google AI Studio

Could you please test with gemini-3.8-flash-tts and let us know if you are still experiencing the same audio issue?

3

I’ll check today whether the same bug is present in Gemini 3.8 Flash TTS, because I’ve encountered the same issue in TTS 2.5 Pro as the user above.

But I’d also like to request a feature for Gemini 3.8 Flash TTS that is already available in 2.5 Pro: the ability to choose between Text and Composer modes.

Right now, Gemini 3.8 only has Composer mode, which means that when I create a dialogue, I have to manually insert each line one by one.

With TTS 2.5 Pro, I could prepare the entire script separately, label the speakers in the text as Speaker 1 and Speaker 2, paste the whole script into TTS, and it would automatically assign the correct voices based on those speaker labels and my instructions.

This is implemented very conveniently in Google TTS 2.5 Pro, and I’d really like to see the same functionality added to Gemini 3.8 Flash TTS.

4

@Sai_Deepika_K
I tested Gemini 3.8 Flash TTS today.

The good news is that I did not encounter the audio glitch that I was experiencing with Gemini 2.5 Pro TTS. I did not hear the random micro-cuts or discontinuities inside words, and the generated audio sounded clean and good overall.

However, I found several workflow limitations in Gemini 3.8 that make it significantly less convenient for generating complete scripted dialogues.

First, Gemini 3.8 seems to respond less conveniently to a single global voice-direction prompt for the entire script compared with Gemini 2.5 Pro. In 2.5 Pro, I could provide one general prompt describing how the whole dialogue should be performed. In 3.8, it appears that I need to add voice or style instructions separately for individual speaker blocks.

Second, as I mentioned above, Gemini 2.5 Pro allowed me to prepare the entire script in advance using labels such as:

Speaker 1:
Speaker 2:

I could paste the complete dialogue at once, assign the voices, and generate it as one script.

In Gemini 3.8, I currently have to manually create the speaker blocks and paste the individual lines into them. For longer dialogues, this adds a lot of unnecessary manual work.

I also noticed something confusing about the Play button above each individual speech block.

At first, I thought this was a very useful feature for regenerating only one problematic line without losing the rest of the already generated dialogue.

For example, I generated a dialogue of about two minutes and noticed that one particular line was not spoken well. I clicked Play on that individual block, and the model generated a new version of that line that sounded better.

However, when I later downloaded the complete audio file, the final file still contained the original version of that line, not the newly generated version.

So at the moment, the purpose of this button is unclear to me. It seems to generate a new preview of the individual line, but that regenerated result is not applied to the final dialogue.

Ideally, the per-block controls should support two separate functions:

  1. Play the exact current version of that line as it exists in the full generated dialogue.
  2. Regenerate only that specific line and replace it in the final dialogue without regenerating or losing all the other already generated lines.

That would be extremely useful for longer scripts, because often only one or two lines need to be corrected.

So, regarding the original audio-quality issue: Gemini 3.8 Flash TTS works well for me so far and I have not reproduced the micro-cut bug from Gemini 2.5 Pro.

The main issues I see in 3.8 are currently related to the script workflow and per-line editing/regeneration rather than audio quality.

5

Hi! Thanks for replying. I’ve been testing Gemini 3.8 TTS for the past two days and I think it’s a substantial improvement in audio quality, both in naturalness and expression.

However, there’s no option to customize it with British voices, which makes it practically useless to me. I’m aware that the voices can be adjusted in the voice audio profile, but when I type something like “London Accent” or “British Accent,” it usually produces a voice that’s a sort of mix between American and British. If you know of any way to fix this, I’d appreciate it if you could let me know.

By the way, Gemini 2.5 Pro TTS still has some glitches and audio micro-skips.

6

Hello @OMAN ,

Thanks for confirming that gemini-3.8-flash-tts resolves the audio glitches, and for the detailed notes on the multi-speaker workflow. They’re really helpful.

7

I have been experiencing this same CD skipping, audio drops, everything mentioned above. It seemed to start the day gemini tts began strong-arming me for more and more tokens, cutting off my small daily projects in the middle of the project (I had plenty of credits). I have been using 2.5 pro for over a year. This month, the cutting off my project and demanding more money at the same exact time this unusable tts started spitting out HORRIBLE voice generations. The other (newer) tts models (as mentioned), are absolutely horrible for the workflow I have been doing for quite some time now. Any idea how long we have to keep paying for garbage until it’s fixed?? Thank you. I appreciate your time.

8

I’m in the same boat. 90% of my Gemini AI Pro plan usage is for TTS, to generate audio for my educational website. I’ve never had any problems with Gemini 2.5 Pro TTS, and I think it’s one of the best models currently available for audio generation. For some reason, and I’m starting to think it was intentional, it’s barely usable right now due to frequent glitches and audio dropouts. It seems there’s been a general decline in quality with this model.

While it’s true that the new Gemini 3.8 model is very natural and capable of generating many vocal nuances, most of the time it sounds excessively metallic. Furthermore, it doesn’t have a native setting to change the accent. I work with a British accent, and the only way to change it is to change the audio profile description to ‘British Accent’. Even then, it comes out as a mix of American and British accents.

9

Seriously… how is it NOT supposed to look intentional?

People here are asking for 2.5 to work correctly and the only answer that keeps getting repeated, “try out new 3.8” …so, you are destroying 2.5 to try and force users into the newest version? That’s pretty shady. YOU STILL TAKE OUR MONEY FOR 2.5!