Audio modelActive

Gemini 3.8 Flash-Lite TTS

Google

Released: September 22, 2026
Data date: October 7, 2026

Model class: Text-to-speech model
Access: Gemini API

Gemini 3.8 Flash-Lite TTS is Google's speech model tuned for throughput and cost efficiency in production workloads and real-time agents.

Gemini 3.8 Flash TTS targets higher voice fidelity, nuanced acting, and regional dialects; Flash-Lite prioritizes scalable generation.

Specifications and access

Specification	Value and source
Model class	Text-to-speech modelSource
Access	Gemini APISource
Input modalities	TextSource
Output modalities	AudioSource
API model ID	gemini-3.8-flash-lite-ttsSource
Focus	Designed for high-throughput, cost-efficient production and real-time voice-agent cascadesSource

Measurements without matching peer values

These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.

Artificial Analysis Provider Voice Arena Elo (Show measurement, test conditions, and source)

Source value: 1242
Score: 1242
Metric: human blind-pairwise voice preference Elo
Unit: elo
Benchmark version: September 2026 snapshot
Category: audio
Direction: Higher is better
Sample: 2372
Confidence interval: 15
Lower confidence bound: 1227
Upper confidence bound: 1257
Source type: independent-evaluator
Evaluator: Artificial Analysis Provider Voice Arena
Status: active
Retrieved at: 2026-10-07
Methodology: Independent evaluation. Blind pairwise evaluation used each provider's native voice set; 2,372 model appearances across eight voices. Higher Elo means more frequent preference. September 2026 snapshot.
Context: Appearances are not a count of unique people or prompts. The snapshot does not give a per-model evaluation date range.

Source Artificial Analysis Provider Voice Arena

Hume Real World VoiceEQ: Extended long-form stability (Show measurement, test conditions, and source)

Source value: 3.03
Score: 3.03
Metric: five-point stability rating
Unit: rating
Benchmark version: Report card dated 2026-09-24
Category: audio
Direction: Higher is better
Source type: independent-evaluator
Evaluator: Hume Real World VoiceEQ report card
Status: active
Retrieved at: 2026-10-07
Methodology: Independent evaluation. Three human raters screened by Hume; blinded evaluation with randomized clip order on held-out prompts. Gemini 3.8 Flash-Lite TTS; long-form stability dimension.
Context: Hume publishes no per-question sample counts, confidence intervals, or model configurations. It documents three screened human raters, blinded scoring, randomized order, and held-out prompts. Hume discloses a non-exclusive licensing agreement with Google.

Source Hume Real World VoiceEQ report card

Hume Real World VoiceEQ: Expressivity–Reliability Frontier Score (Show measurement, test conditions, and source)

Source value: 0.914
Score: 0.914
Metric: TOPSIS composite on a 0–1 scale
Unit: ratio
Benchmark version: Report card dated 2026-09-24
Category: audio
Direction: Higher is better
Source type: independent-evaluator
Evaluator: Hume Real World VoiceEQ report card
Status: active
Retrieved at: 2026-10-07
Methodology: Independent evaluation. Three human raters screened by Hume; blinded evaluation with randomized clip order on held-out prompts. Gemini 3.8 Flash-Lite TTS; TOPSIS closeness to the ideal (1, 1) corner across expressivity and reliability questions.
Context: Hume publishes no per-question sample counts, confidence intervals, or model configurations. It documents three screened human raters, blinded scoring, randomized order, and held-out prompts. Hume discloses a non-exclusive licensing agreement with Google.

Source Hume Real World VoiceEQ report card

Hume Real World VoiceEQ: Same-speaker similarity (Show measurement, test conditions, and source)

Source value: 3.68
Score: 3.68
Metric: same-speaker similarity rating out of five
Unit: rating
Benchmark version: Report card dated 2026-09-24
Category: audio
Direction: Higher is better
Source type: independent-evaluator
Evaluator: Hume Real World VoiceEQ report card
Status: active
Retrieved at: 2026-10-07
Methodology: Independent evaluation. Three human raters screened by Hume; blinded evaluation with randomized clip order on held-out prompts. Gemini 3.8 Flash-Lite TTS; voice replication same-speaker dimension; report card ranks this model 7th of 13.
Context: Hume publishes no per-question sample counts, confidence intervals, or model configurations. It documents three screened human raters, blinded scoring, randomized order, and held-out prompts. Hume discloses a non-exclusive licensing agreement with Google.

Source Hume Real World VoiceEQ report card

Hume Real World VoiceEQ: Acting / Role-fit (Show measurement, test conditions, and source)

Source value: 3.65
Score: 3.65
Metric: five-point human-rater score
Unit: rating
Benchmark version: Hume Real World VoiceEQ report card, 2026-09-24
Category: audio
Direction: Higher is better
Source type: independent-evaluator
Evaluator: Hume Real World VoiceEQ report card
Status: active
Retrieved at: 2026-10-07
Methodology: Independent evaluation. Gemini 3.8 Flash-Lite TTS preview; Hume Real World VoiceEQ blind evaluation with randomized clips, three screened human raters, and held-out prompts; Acting / Role-fit dimension, reported cohort of 34 models.
Context: Hume publishes no per-dimension clip counts or confidence intervals. It documents three screened human raters, blinded scoring, randomized clip order, and held-out prompts. Hume discloses a non-exclusive licensing agreement with Google.

Source Hume Real World VoiceEQ report card

Hume Real World VoiceEQ: Expressiveness (Show measurement, test conditions, and source)

Source value: 3.36
Score: 3.36
Metric: five-point human-rater score
Unit: rating
Benchmark version: Hume Real World VoiceEQ report card, 2026-09-24
Category: audio
Direction: Higher is better
Source type: independent-evaluator
Evaluator: Hume Real World VoiceEQ report card
Status: active
Retrieved at: 2026-10-07
Methodology: Independent evaluation. Gemini 3.8 Flash-Lite TTS preview; Hume Real World VoiceEQ blind evaluation with randomized clips, three screened human raters, and held-out prompts; Expressiveness among instruction-capable models dimension, reported cohort of 8 models.
Context: Hume publishes no per-dimension clip counts or confidence intervals. It documents three screened human raters, blinded scoring, randomized clip order, and held-out prompts. Hume discloses a non-exclusive licensing agreement with Google.

Source Hume Real World VoiceEQ report card

Hume Real World VoiceEQ: Voice identity (Show measurement, test conditions, and source)

Source value: 4.61
Score: 4.61
Metric: five-point human-rater score
Unit: rating
Benchmark version: Hume Real World VoiceEQ report card, 2026-09-24
Category: audio
Direction: Higher is better
Source type: independent-evaluator
Evaluator: Hume Real World VoiceEQ report card
Status: active
Retrieved at: 2026-10-07
Methodology: Independent evaluation. Gemini 3.8 Flash-Lite TTS preview; Hume Real World VoiceEQ blind evaluation with randomized clips, three screened human raters, and held-out prompts; Voice identity among instruction-capable models dimension, reported cohort of 8 models.
Context: Hume publishes no per-dimension clip counts or confidence intervals. It documents three screened human raters, blinded scoring, randomized clip order, and held-out prompts. Hume discloses a non-exclusive licensing agreement with Google.

Source Hume Real World VoiceEQ report card

Hume Real World VoiceEQ: Language stability (Show measurement, test conditions, and source)

Source value: 3.17
Score: 3.17
Metric: five-point human-rater score
Unit: rating
Benchmark version: Hume Real World VoiceEQ report card, 2026-09-24
Category: audio
Direction: Higher is better
Source type: independent-evaluator
Evaluator: Hume Real World VoiceEQ report card
Status: active
Retrieved at: 2026-10-07
Methodology: Independent evaluation. Gemini 3.8 Flash-Lite TTS preview; Hume Real World VoiceEQ blind evaluation with randomized clips, three screened human raters, and held-out prompts; Language stability dimension, reported cohort of 34 models.
Context: Hume publishes no per-dimension clip counts or confidence intervals. It documents three screened human raters, blinded scoring, randomized clip order, and held-out prompts. Hume discloses a non-exclusive licensing agreement with Google.

Source Hume Real World VoiceEQ report card

Hume Real World VoiceEQ: Reliability (Show measurement, test conditions, and source)

Source value: 3.6
Score: 3.6
Metric: five-point human-rater score
Unit: rating
Benchmark version: Hume Real World VoiceEQ report card, 2026-09-24
Category: audio
Direction: Higher is better
Source type: independent-evaluator
Evaluator: Hume Real World VoiceEQ report card
Status: active
Retrieved at: 2026-10-07
Methodology: Independent evaluation. Gemini 3.8 Flash-Lite TTS preview; Hume Real World VoiceEQ blind evaluation with randomized clips, three screened human raters, and held-out prompts; Reliability dimension, reported cohort of 34 models.
Context: Hume publishes no per-dimension clip counts or confidence intervals. It documents three screened human raters, blinded scoring, randomized clip order, and held-out prompts. Hume discloses a non-exclusive licensing agreement with Google.

Source Hume Real World VoiceEQ report card

Hume Real World VoiceEQ: Long-form speaker stability (Show measurement, test conditions, and source)

Source value: 3.4
Score: 3.4
Metric: five-point human-rater score
Unit: rating
Benchmark version: Hume Real World VoiceEQ report card, 2026-09-24
Category: audio
Direction: Higher is better
Source type: independent-evaluator
Evaluator: Hume Real World VoiceEQ report card
Status: active
Retrieved at: 2026-10-07
Methodology: Independent evaluation. Gemini 3.8 Flash-Lite TTS preview; Hume Real World VoiceEQ blind evaluation with randomized clips, three screened human raters, and held-out prompts; Long-form speaker stability dimension, reported cohort of 34 models.
Context: Hume publishes no per-dimension clip counts or confidence intervals. It documents three screened human raters, blinded scoring, randomized clip order, and held-out prompts. Hume discloses a non-exclusive licensing agreement with Google.

Source Hume Real World VoiceEQ report card

Hume Real World VoiceEQ: Multilingual long-form speaker stability (Show measurement, test conditions, and source)

Source value: 3.3
Score: 3.3
Metric: five-point human-rater score
Unit: rating
Benchmark version: Hume Real World VoiceEQ report card, 2026-09-24
Category: audio
Direction: Higher is better
Source type: independent-evaluator
Evaluator: Hume Real World VoiceEQ report card
Status: active
Retrieved at: 2026-10-07
Methodology: Independent evaluation. Gemini 3.8 Flash-Lite TTS preview; Hume Real World VoiceEQ blind evaluation with randomized clips, three screened human raters, and held-out prompts; Multilingual long-form speaker stability dimension, reported cohort of 24 models.
Context: Hume publishes no per-dimension clip counts or confidence intervals. It documents three screened human raters, blinded scoring, randomized clip order, and held-out prompts. Hume discloses a non-exclusive licensing agreement with Google.

Source Hume Real World VoiceEQ report card

Hume Real World VoiceEQ: Acoustic quality (Show measurement, test conditions, and source)

Source value: 3.24
Score: 3.24
Metric: five-point human-rater score
Unit: rating
Benchmark version: Hume Real World VoiceEQ report card, 2026-09-24
Category: audio
Direction: Higher is better
Source type: independent-evaluator
Evaluator: Hume Real World VoiceEQ report card
Status: active
Retrieved at: 2026-10-07
Methodology: Independent evaluation. Gemini 3.8 Flash-Lite TTS preview; Hume Real World VoiceEQ blind evaluation with randomized clips, three screened human raters, and held-out prompts; Acoustic quality dimension, reported cohort of 34 models.
Context: Hume publishes no per-dimension clip counts or confidence intervals. It documents three screened human raters, blinded scoring, randomized clip order, and held-out prompts. Hume discloses a non-exclusive licensing agreement with Google.

Source Hume Real World VoiceEQ report card

Sources and data date

Type	Evidence and data date
Research status	Research date October 7, 2026. 4 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Not published	hume-sample-counts-and-intervals Original wording Hume publishes no per-model clip or rating counts and no confidence intervals for the cited VoiceEQ metrics.
Not found in the inspected sources	objective-speech-error-metrics Original wording The sources checked contain no transcript-based WER/CER test for the generated TTS audio; available results measure voice quality, stability, and preference.
Additional source	Google Gemini API release notes (retrieved October 7, 2026) · Google Gemini API release notes · Editorial description reviewed October 7, 2026
Additional source	Google Gemini text-to-speech documentation (retrieved October 7, 2026) · Google Gemini text-to-speech documentation · Editorial description reviewed October 7, 2026
Additional source	Gemini 3.8 Flash-Lite TTS model documentation (retrieved October 7, 2026)
Additional source	Artificial Analysis Provider Voice Arena (retrieved October 7, 2026)
Additional source	Hume Real World VoiceEQ report card (retrieved October 7, 2026)