Audio modelActive
Eleven v4 Turbo
ElevenLabs
- Released
- September 28, 2026
- Data date
- October 7, 2026
- Model class
- Languages
Eleven v4 Turbo brings v4's expressive speech generation to real-time applications and voice agents. ElevenLabs optimized the variant alongside its Agents platform.
While Eleven v4 is recommended for high-fidelity content such as audiobooks, Turbo is the v4 option for live conversations; both models support more than 90 languages.
Specifications and access
| Specification | Value and source |
|---|---|
| Model class | Text-to-speech modelSource |
| Access | ElevenAPI and ElevenLabs productsSource |
| Input modalities | TextSource |
| Output modalities | AudioSource |
| API model ID | eleven_v4_turboSource |
| Languages | More than 90Source |
| Focus | Expressive speech generation with low latency for real-time applications and voice agentsSource |
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.
Coval TTS Prompt Set speech intelligibility WER (Show measurement, test conditions, and source)
- Source value
- 1.8
- Score
- 1.8
- Metric
- WER (TTS intelligibility word error rate; percent)
- Unit
- %
- Benchmark version
- Coval public TTS benchmark, rolling 30-day window read 2026-10-07
- Category
- audio
- Direction
- Lower is better
- Sample
- 566
- Source type
- independent-evaluator
- Evaluator
- Coval Voice AI benchmark, Eleven v4 Turbo endpoint
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. Eleven v4 Turbo served through the official ElevenLabs API on Coval fixed TTS text prompts. Coval transcribes generated audio with fixed OpenAI whisper-1 and compares the transcript with the input text; this is the TTS Prompt Set rolling 30-day WER.
- Context
- This is an intelligibility proxy from a fixed ASR model, not an ASR benchmark of Eleven v4 Turbo. Coval notes that whisper-1 errors affect every TTS provider. The rolling-window count is specific to this WER metric and differs from the TTFA count.
Coval TTS Prompt Set speech intelligibility WER p95 (Show measurement, test conditions, and source)
- Source value
- 13
- Score
- 13
- Metric
- WER 95th percentile (TTS intelligibility; percent)
- Unit
- %
- Benchmark version
- Coval public TTS benchmark, rolling 30-day window read 2026-10-07
- Category
- audio
- Direction
- Lower is better
- Sample
- 566
- Source type
- independent-evaluator
- Evaluator
- Coval Voice AI benchmark, Eleven v4 Turbo endpoint
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. Eleven v4 Turbo served through the official ElevenLabs API on Coval fixed TTS text prompts. Coval transcribes generated audio with fixed OpenAI whisper-1 and compares the transcript with the input text; this is the TTS Prompt Set rolling 30-day WER distribution, 95th percentile.
- Context
- This is an intelligibility proxy from a fixed ASR model, not an ASR benchmark of Eleven v4 Turbo. Coval notes that whisper-1 errors affect every TTS provider. The rolling-window count is specific to this WER metric and differs from the TTFA count.
Artificial Analysis Provider Voice Arena Elo (Show measurement, test conditions, and source)
- Source value
- 1334
- Score
- 1334
- Metric
- human blind-pairwise voice preference Elo
- Unit
- elo
- Benchmark version
- September 2026 snapshot
- Category
- audio
- Direction
- Higher is better
- Sample
- 1645
- Confidence interval
- 19
- Lower confidence bound
- 1315
- Upper confidence bound
- 1353
- 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; 1,645 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.
ElevenLabs approximate median inference latency (Show measurement, test conditions, and source)
- Source value
- 0.1
- Score
- 0.1
- Metric
- median inference latency
- Unit
- seconds
- Benchmark version
- ElevenLabs measurement reported September 2026
- Category
- audio
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- ElevenLabs Eleven v4 Turbo launch post
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Eleven v4 Turbo in ElevenAgents. The launch post reports an approximate median inference latency of 100 ms. The timing boundaries and measurement setup are not specified.
- Context
- This is an approximate provider claim. The source does not define timing boundaries or publish run count, hardware, spread or confidence intervals. The separate Eleven v4 listener-preference setup does not document this latency measurement.
ElevenLabs approximate median time to first speech in ElevenAgents (Show measurement, test conditions, and source)
- Source value
- 0.15
- Score
- 0.15
- Metric
- median time from request to audible speech
- Unit
- seconds
- Benchmark version
- ElevenLabs launch measurement, September 2026
- Category
- audio
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- ElevenLabs Eleven v4 launch report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Eleven v4 Turbo in the vendor-reported TTS measurement; median request-to-audible-speech time with identical scripts and default settings compared against Cartesia Sonic 3.6, xAI TTS, Gemini 3.8 Flash-Lite TTS, and GPT-4o mini TTS. Eleven v4 Turbo used WebSocket streaming and the measurement removed network latency.
- Context
- ElevenLabs reports an approximate value but publishes no sample count or confidence interval. The vendor describes Eleven v4 Turbo and ElevenAgents as a co-optimized system; this is not independent evidence or a full caller-to-agent end-to-end latency measurement. Do not compare directly with endpoint-level TTFA without preserving the measurement boundary.
Coval text-to-speech Time to First Audio mean (Show measurement, test conditions, and source)
- Source value
- 0.204
- Score
- 0.204
- Metric
- TTFA arithmetic mean
- Unit
- seconds
- Benchmark version
- Coval public TTS benchmark, rolling 30-day window read 2026-10-07
- Category
- audio
- Direction
- Lower is better
- Sample
- 568
- Source type
- independent-evaluator
- Evaluator
- Coval Voice AI benchmark, Eleven v4 Turbo endpoint
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. Eleven v4 Turbo served through the official ElevenLabs API and measured by Coval with the same fixed TTS text prompts; Coval runner in us-east-1. TTFA runs from synthesis request to first audible sample; TCP/TLS/handshake setup is excluded, while leading silence and the round trip to the provider region are included. This is the published arithmetic mean across the rolling 30-day sample.
- Context
- Independent hosted-endpoint measurement over a rolling 30-day window, re-measured daily; the sample count is the page value read on 2026-10-07 and will change. It is API/system timing, not standalone model inference; serving region, voice, prompt, format, and network path affect the result.
Coval text-to-speech Time to First Audio p50 (Show measurement, test conditions, and source)
- Source value
- 0.194
- Score
- 0.194
- Metric
- TTFA 50th percentile
- Unit
- seconds
- Benchmark version
- Coval public TTS benchmark, rolling 30-day window read 2026-10-07
- Category
- audio
- Direction
- Lower is better
- Sample
- 568
- Source type
- independent-evaluator
- Evaluator
- Coval Voice AI benchmark, Eleven v4 Turbo endpoint
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. Eleven v4 Turbo served through the official ElevenLabs API and measured by Coval with the same fixed TTS text prompts; Coval runner in us-east-1. TTFA runs from synthesis request to first audible sample; TCP/TLS/handshake setup is excluded, while leading silence and the round trip to the provider region are included. This is the published 50th percentile across the rolling 30-day sample.
- Context
- Independent hosted-endpoint measurement over a rolling 30-day window, re-measured daily; the sample count is the page value read on 2026-10-07 and will change. It is API/system timing, not standalone model inference; serving region, voice, prompt, format, and network path affect the result.
Coval text-to-speech Time to First Audio p99 (Show measurement, test conditions, and source)
- Source value
- 0.445
- Score
- 0.445
- Metric
- TTFA 99th percentile
- Unit
- seconds
- Benchmark version
- Coval public TTS benchmark, rolling 30-day window read 2026-10-07
- Category
- audio
- Direction
- Lower is better
- Sample
- 568
- Source type
- independent-evaluator
- Evaluator
- Coval Voice AI benchmark, Eleven v4 Turbo endpoint
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. Eleven v4 Turbo served through the official ElevenLabs API and measured by Coval with the same fixed TTS text prompts; Coval runner in us-east-1. TTFA runs from synthesis request to first audible sample; TCP/TLS/handshake setup is excluded, while leading silence and the round trip to the provider region are included. This is the published 99th percentile across the rolling 30-day sample.
- Context
- Independent hosted-endpoint measurement over a rolling 30-day window, re-measured daily; the sample count is the page value read on 2026-10-07 and will change. It is API/system timing, not standalone model inference; serving region, voice, prompt, format, and network path affect the result.
Coval text-to-speech TTFA component: network roundtrip (Show measurement, test conditions, and source)
- Source value
- 0.143
- Score
- 0.143
- Metric
- TTFA component: network roundtrip
- Unit
- seconds
- Benchmark version
- Coval public TTS benchmark, rolling 30-day window read 2026-10-07
- Category
- audio
- Direction
- Lower is better
- Sample
- 568
- Source type
- independent-evaluator
- Evaluator
- Coval Voice AI benchmark, Eleven v4 Turbo endpoint
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. Eleven v4 Turbo served through the official ElevenLabs API and measured by Coval with the same fixed TTS text prompts; Coval runner in us-east-1. TTFA runs from synthesis request to first audible sample; TCP/TLS/handshake setup is excluded, while leading silence and the round trip to the provider region are included. Coval publishes this component as the mean network round trip in its TTFA decomposition.
- Context
- Independent hosted-endpoint measurement over a rolling 30-day window, re-measured daily; the sample count is the page value read on 2026-10-07 and will change. It is API/system timing, not standalone model inference; serving region, voice, prompt, format, and network path affect the result.
Coval text-to-speech TTFA component: leading silence (Show measurement, test conditions, and source)
- Source value
- 0.061
- Score
- 0.061
- Metric
- TTFA component: leading silence
- Unit
- seconds
- Benchmark version
- Coval public TTS benchmark, rolling 30-day window read 2026-10-07
- Category
- audio
- Direction
- Lower is better
- Sample
- 568
- Source type
- independent-evaluator
- Evaluator
- Coval Voice AI benchmark, Eleven v4 Turbo endpoint
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. Eleven v4 Turbo served through the official ElevenLabs API and measured by Coval with the same fixed TTS text prompts; Coval runner in us-east-1. TTFA runs from synthesis request to first audible sample; TCP/TLS/handshake setup is excluded, while leading silence and the round trip to the provider region are included. Coval publishes this component as the mean leading silence before audible speech in its TTFA decomposition.
- Context
- Independent hosted-endpoint measurement over a rolling 30-day window, re-measured daily; the sample count is the page value read on 2026-10-07 and will change. It is API/system timing, not standalone model inference; serving region, voice, prompt, format, and network path affect the result.
Coval text-to-speech Time to First Audio p95 (Show measurement, test conditions, and source)
- Source value
- 0.272
- Score
- 0.272
- Metric
- TTFA 95th percentile
- Unit
- seconds
- Benchmark version
- Coval public TTS benchmark, rolling 30-day window read 2026-10-07
- Category
- audio
- Direction
- Lower is better
- Sample
- 568
- Source type
- independent-evaluator
- Evaluator
- Coval Voice AI benchmark, Eleven v4 Turbo endpoint
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. Eleven v4 Turbo served through the official ElevenLabs API and measured by Coval with the same fixed TTS text prompts; Coval runner in us-east-1. TTFA runs from synthesis request to first audible sample; TCP/TLS/handshake setup is excluded, while leading silence and the round trip to the provider region are included. This is the published 95th percentile across the rolling 30-day sample.
- Context
- Independent hosted-endpoint measurement over a rolling 30-day window, re-measured daily; the sample count is the page value read on 2026-10-07 and will change. It is API/system timing, not standalone model inference; serving region, voice, prompt, format, and network path affect the result.