# Gemini 3.1 Flash TTS review: cheap and expressive, but broken past a minute

> Google's Gemini 3.1 Flash TTS lands mid-table on Artificial Analysis and delivers strong short-form output at a low price, but a confirmed long-form voice-drift defect and no voice cloning cap its usefulness.

## Key Information
- Published: Sep 10, 2026
- Category: Reviews
- Summary: Gemini 3.1 Flash TTS is one of the cheapest, most prompt-steerable text-to-speech models available, and for clips under a minute it is genuinely good. Past that, it has a confirmed, still-unfixed defect that makes it unsuitable for audiobooks or long narration.

## Content
## Verdict

Gemini 3.1 Flash TTS earns its place in the middle of the texttospeech.com aggregate, and that placement is honest: it is a strong short-form model with a known long-form failure. It is cheap ($1 per million input tokens, $20 per million output on OpenRouter), expressive, prompt-steerable, and covers 70+ languages. But it degrades past roughly 60 seconds of generated audio — voice drift and a volume/quality drop that Google staff have acknowledged and are still investigating as of August 2026 — and it has no voice cloning, which keeps it out of Vapi's Humanness Index entirely.

For short demos, explainer audio, and experimentation, it is a good default. For long narration, it is not ready.

## What it is

Gemini 3.1 Flash TTS is Google's native text-to-speech model, launched around April 15, 2026. It generates audio directly rather than through a separate TTS pipeline, which cuts the moving parts down to a single API call. It outputs 24kHz PCM, supports multi-speaker audio and audio tags in its output, and speaks more than 70 languages.

The launch was notable enough that independent developer Simon Willison covered it the same day, calling the required prompting approach "surprising" — the model wants theatrical, director's-notes style prompts to get the delivery right ([simonwillison.net, Apr 15 2026](https://simonwillison.net/2026/Apr/15/gemini-31-flash-tts/)).

## Where it sits across the boards

The texttospeech.com aggregate is the output of a published methodology, not a unanimous verdict — it folds three public blind-test boards into a single 0–100 score. Gemini 3.1 Flash TTS's position differs board to board, and the honest read is mixed.

**Artificial Analysis Speech Arena** — the strongest, most prominent signal. Gemini 3.1 Flash TTS currently sits around **1,210 Quality Elo**, roughly **8th**, having slipped from a launch position of #2 at ~1,211 Elo as the field has grown ([Artificial Analysis leaderboard](https://artificialanalysis.ai/text-to-speech/leaderboard); [model page](https://artificialanalysis.ai/text-to-speech/models/gemini-3-1-tts)). The drop is not evidence the model got worse — the field grew around it. But it is no longer a top-two contender on the blind-test Elo that matters most.

**Vapi Humanness Index** — Gemini 3.1 Flash TTS is **absent**, and that absence is itself a data point. The index only admits models that support voice cloning; Gemini offers none, so it is ineligible rather than poorly ranked ([Vapi Humanness Index](https://humannessindex.vapi.ai/)). On humanness, the model has effectively no score because it cannot be tested the way the field's leaders can.

**EmergentTTS-Eval** — the NeurIPS 2025 benchmark scores the family, not this generation: Gemini-2.5-Flash-Preview-TTS placed #1 overall with a 75.57% win-rate and 10.39 WER ([arXiv:2505.23009](https://arxiv.org/abs/2505.23009); [public repo](https://github.com/boson-ai/EmergentTTS-Eval-public)). That is a baseline for the family's synthetic speech quality, not a score for 3.1 itself, but it shows the lineage is strong.

## What wider independent evidence shows

The confirmed long-form defect is the story here, and it is well documented.

A thread on the Google AI Developers Forum — user-reported and acknowledged by Google — describes volume and quality dropping and voice drift on TTS requests longer than about one minute, plus "runaway" PCM errors and TTS breaking when temperature is set below 0.6 ([forum thread](https://discuss.ai.google.dev/t/gemini-3-1-flash-live-voice-slowly-changing-massive-audio-quality-volume-dropping-on-tts-requests-longer-than-1-minute/142499)). Google staff confirmed the issue on August 5, 2026 and said engineering was investigating, with no fix or timeline.

Independent hands-on testing from TTSAudit, a smaller outlet, quantifies the same thing: roughly 90% of generations over one minute degrade, and a three-minute sample scored 92 on its opening 16 seconds but 34 on its final 20 ([TTSAudit](https://ttsaudit.com/blog/gemini-3-1-flash-tts-long-form-quality)). The same review praises the sub-60-second output — expressive, with audio tags and multi-speaker support across 70+ languages — which matches the broader consensus: strong short, broken long.

On the Hacker News thread at launch, one point worth flagging for anyone relying on the prompting guidance: voice is set by the voice parameter, and the prompt only directs *how the chosen voice delivers* ([HN thread, 18 points](https://news.ycombinator.com/item?id=47782610)). Google Developer Expert Evan Lin, writing on dev.to, confirms the practical angle — native 24kHz and roughly 30% less code, but Traditional-Chinese pronunciation that only comes right through prompt engineering ([dev.to](https://dev.to/gde/gemini-31-native-tts-for-easier-more-powerful-summary-reading-2ep9)). Both are affiliation-flagged, so weigh accordingly.

Platform telemetry rounds out the operational picture. On OpenRouter, Gemini 3.1 Flash TTS preview shows **1.50s P50 latency**, 100% 3-day uptime with 97.27% availability, at $1/1M input and $20/1M output ([OpenRouter](https://openrouter.ai/google/gemini-3.1-flash-tts-preview)). A MarkTechPost TTFT benchmark from August 30, 2026 lists it at Elo 1,210, ~$18.30/1M, 6th ([MarkTechPost](https://www.marktechpost.com/2026/08/30/lowest-latency-inference-apis-for-voice-and-realtime-agents-a-time-to-first-token-ttft-first-benchmark/)).

## Strengths

- **Price.** At $1/1M input and $20/1M output on OpenRouter, it undercuts most serious competitors. The cheapest credible option in its class.
- **Prompt-steerability.** Director's-notes prompting gives fine control over delivery — the model is more steerable than most TTS endpoints.
- **Languages.** 70+ languages is real coverage, not a marketing line.
- **Output features.** Multi-speaker audio and audio tags in a single generation.
- **Simplicity.** Native TTS means no long-lived WebSocket, roughly 30% less integration code, and 24kHz PCM output.

## Weaknesses

- **Long-form degradation.** The confirmed defect is the headline problem. Past ~60 seconds, voice drift and volume/quality drop appear, and Google has not shipped a fix. This rules out audiobooks and long narration outright.
- **No voice cloning.** This is a capability gap, not a quality gap, but it has a concrete cost: exclusion from Vapi's Humanness Index, and no path to cloned-voice parity with the field's leaders.
- **Rank slippage.** From #2 at launch to ~8th on Artificial Analysis as the field grew. Not a decline in the model — a sign of how fast the top of the field is moving.
- **Prompting friction.** The required theatrical prompt style is unusual, and per the HN thread, voice identity is set by the voice parameter — the prompt only shapes delivery. Easy to expect more control than you get.
- **Fragility.** Temperature below 0.6 breaks TTS, and PCM "runaway" errors appear in long requests.

## Who it's for

**For:** developers and content teams generating short clips — demos, explainer audio, product walkthroughs, multi-speaker scenes, or anything under a minute in one of the 70+ supported languages. Anyone who wants cheap, expressive, steerable TTS and can live within the length limit. Experimenters testing what prompt-steerable native TTS can do.

**Not for:** audiobook and long-form narration pipelines, voice-cloning workflows, or production systems that need a fixed, consistent voice across minutes of audio. Any use case where a voice that drifts past 60 seconds is a defect, not a tradeoff. If humanness-score parity with cloned-voice models matters to you, this model is not in the running.

## Bottom line

Gemini 3.1 Flash TTS is the best cheap short-form TTS on the market right now, and it is not close on price-to-quality for sub-60-second audio. The expressive delivery, 70+ languages, and prompt control are real. But it sits mid-table on the aggregate for a reason: a confirmed, still-unfixed long-form defect, no voice cloning, and a field that has grown past its launch position.

Use it for short clips and experimentation. Wait for a fix before putting it anywhere near long narration.

## Links
- Read full article: https://texttospeech.com/blog/gemini-3-1-flash-tts-review
- All posts: https://texttospeech.com/blog
- Leaderboard: https://texttospeech.com/