Audio modelActive
Muse Voice Transcribe
Meta Superintelligence Labs
- Released
- September 1, 2026
- Data date
- October 7, 2026
- Languages
Muse Voice Transcribe handles transcription, speaker diarization, and endpoint detection inside one streaming model. Meta lists more than 20 distinguishable speakers and language switching within a conversation.
The model adjusts how long it waits before emitting a word according to its difficulty. Meta describes this as a way to balance accuracy against delay; audio input longer than one hour is also supported.
Specifications and access
| Specification | Value and source |
|---|---|
| Model class | Streaming speech recognition with diarizationSource |
| Access | Meta Model API, Meta AI for Mac, Muse Code, and browser demoSource |
| Input modalities | Live audioSource |
| Output modalities | Text transcriptSource |
| API model ID | muse-voice-transcribe-1.0Source |
| Languages | 25+; supports code-switchingSource |
| Speaker diarization | Real-time diarization for more than 20 speakersSource |
| Long audio input | Meta specifies native audio inputs longer than one hourSource |
| Published access | Available through Meta Model API, Meta AI for Mac, and Muse Code, with an additional browser microphone demo.Source |
Page 1 of 2
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.
Meta average diarization error rate (Show measurement, test conditions, and source)
- Source value
- 17.5
- Score
- 17.5
- Metric
- average diarization error rate
- Unit
- %
- Benchmark version
- Meta announcement chart, September 2026
- Category
- audio
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- Meta Muse Voice Transcribe announcement chart
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Average across AMI-IHM, AMI-SDM, and VoxConverse. Meta's announcement chart does not give sample counts per dataset.
- Context
- Vendor chart; no per-dataset values, sample counts, or confidence intervals are published.
Artificial Analysis Overall AA-WER Streaming WER (Show measurement, test conditions, and source)
- Source value
- 0.030623
- Score
- 0.030623
- Metric
- word error rate; source's raw fraction
- Unit
- ratio
- Benchmark version
- AA-WER v2.2 streaming; retrieved 2026-10-07
- Category
- audio
- Direction
- Lower is better
- Sample
- 1103
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis AA-WER Streaming
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. Exact Muse Voice Transcribe 1.0 row; AA-AgentTalk 50%, VoxPopuli-Cleaned-AA 25%, Earnings22-Cleaned-AA 25%; approximately eight hours total. WER is duration-weighted within datasets. The overall row uses the published dataset weighting.
- Context
- AA tests the API with English-language datasets; the approximately eight hours do not establish multilingual coverage. Network and streaming configuration are part of the measured endpoint.
Artificial Analysis AA-AgentTalk WER (Show measurement, test conditions, and source)
- Source value
- 0.027419
- Score
- 0.027419
- Metric
- word error rate; source's raw fraction
- Unit
- ratio
- Benchmark version
- AA-WER v2.2 streaming; retrieved 2026-10-07
- Category
- audio
- Direction
- Lower is better
- Sample
- 469
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis AA-WER Streaming
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. Exact Muse Voice Transcribe 1.0 row; proprietary held-out voice-agent speech WER is duration-weighted within datasets. The overall row uses the published dataset weighting.
- Context
- AA tests the API with English-language datasets; the approximately eight hours do not establish multilingual coverage. Network and streaming configuration are part of the measured endpoint.
Artificial Analysis VoxPopuli-Cleaned-AA WER (Show measurement, test conditions, and source)
- Source value
- 0.024583
- Score
- 0.024583
- Metric
- word error rate; source's raw fraction
- Unit
- ratio
- Benchmark version
- AA-WER v2.2 streaming; retrieved 2026-10-07
- Category
- audio
- Direction
- Lower is better
- Sample
- 628
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis AA-WER Streaming
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. Exact Muse Voice Transcribe 1.0 row; English parliamentary speech WER is duration-weighted within datasets. The overall row uses the published dataset weighting.
- Context
- AA tests the API with English-language datasets; the approximately eight hours do not establish multilingual coverage. Network and streaming configuration are part of the measured endpoint.
Artificial Analysis Earnings22-Cleaned-AA WER (Show measurement, test conditions, and source)
- Source value
- 0.043074
- Score
- 0.043074
- Metric
- word error rate; source's raw fraction
- Unit
- ratio
- Benchmark version
- AA-WER v2.2 streaming; retrieved 2026-10-07
- Category
- audio
- Direction
- Lower is better
- Sample
- 6
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis AA-WER Streaming
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. Exact Muse Voice Transcribe 1.0 row; six parent earnings-call files, chunked for streaming as documented WER is duration-weighted within datasets. The overall row uses the published dataset weighting.
- Context
- AA tests the API with English-language datasets; the approximately eight hours do not establish multilingual coverage. Network and streaming configuration are part of the measured endpoint.
Artificial Analysis Overall AA-WER Streaming, first partial after endpoint WER (Show measurement, test conditions, and source)
- Source value
- 0.03574691459283201
- Score
- 0.03574691459283201
- Metric
- word error rate of first post-endpoint partial; raw fraction
- Unit
- ratio
- Benchmark version
- AA-WER v2.2 streaming; retrieved 2026-10-07
- Category
- audio
- Direction
- Lower is better
- Sample
- 1103
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis AA-WER Streaming
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. Exact Muse Voice Transcribe 1.0 row; AA-AgentTalk 50%, VoxPopuli-Cleaned-AA 25%, Earnings22-Cleaned-AA 25%. First partial after detected end of speech; AA reports it separately from final-transcript WER.
- Context
- AA says the first partial is not directly comparable with the later final transcript; the metrics represent different output times.
Artificial Analysis AA-AgentTalk, first partial after endpoint WER (Show measurement, test conditions, and source)
- Source value
- 0.032352118995203585
- Score
- 0.032352118995203585
- Metric
- word error rate of first post-endpoint partial; raw fraction
- Unit
- ratio
- Benchmark version
- AA-WER v2.2 streaming; retrieved 2026-10-07
- Category
- audio
- Direction
- Lower is better
- Sample
- 469
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis AA-WER Streaming
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. Exact Muse Voice Transcribe 1.0 row; proprietary held-out voice-agent speech First partial after detected end of speech; AA reports it separately from final-transcript WER.
- Context
- AA says the first partial is not directly comparable with the later final transcript; the metrics represent different output times.
Artificial Analysis VoxPopuli-Cleaned-AA, first partial after endpoint WER (Show measurement, test conditions, and source)
- Source value
- 0.035094483543360006
- Score
- 0.035094483543360006
- Metric
- word error rate of first post-endpoint partial; raw fraction
- Unit
- ratio
- Benchmark version
- AA-WER v2.2 streaming; retrieved 2026-10-07
- Category
- audio
- Direction
- Lower is better
- Sample
- 628
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis AA-WER Streaming
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. Exact Muse Voice Transcribe 1.0 row; English parliamentary speech First partial after detected end of speech; AA reports it separately from final-transcript WER.
- Context
- AA says the first partial is not directly comparable with the later final transcript; the metrics represent different output times.
Artificial Analysis Earnings22-Cleaned-AA, first partial after endpoint WER (Show measurement, test conditions, and source)
- Source value
- 0.043188936837560865
- Score
- 0.043188936837560865
- Metric
- word error rate of first post-endpoint partial; raw fraction
- Unit
- ratio
- Benchmark version
- AA-WER v2.2 streaming; retrieved 2026-10-07
- Category
- audio
- Direction
- Lower is better
- Sample
- 6
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis AA-WER Streaming
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. Exact Muse Voice Transcribe 1.0 row; six parent earnings-call files, chunked for streaming as documented First partial after detected end of speech; AA reports it separately from final-transcript WER.
- Context
- AA says the first partial is not directly comparable with the later final transcript; the metrics represent different output times.
Artificial Analysis Overall time to final transcript after endpoint (Show measurement, test conditions, and source)
- Source value
- 0.16305052059344063
- Score
- 0.16305052059344063
- Metric
- latency after detected end of speech
- Unit
- seconds
- Benchmark version
- AA-WER v2.2 streaming; retrieved 2026-10-07
- Category
- audio
- Direction
- Lower is better
- Sample
- 1103
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis AA-WER Streaming
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. Exact Muse Voice Transcribe 1.0 row; sample-weighted latency across the AA-WER streaming corpus Timer starts after the speech-end audio was actually streamed and the endpoint signal sent. Network delay remains included.
- Context
- Latency is sample-weighted within datasets; Earnings22 chunks are aggregated back to parent files.
Artificial Analysis AA-AgentTalk time to final transcript after endpoint (Show measurement, test conditions, and source)
- Source value
- 0.16904730040848576
- Score
- 0.16904730040848576
- Metric
- latency after detected end of speech
- Unit
- seconds
- Benchmark version
- AA-WER v2.2 streaming; retrieved 2026-10-07
- Category
- audio
- Direction
- Lower is better
- Sample
- 469
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis AA-WER Streaming
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. Exact Muse Voice Transcribe 1.0 row; proprietary held-out voice-agent speech Timer starts after the speech-end audio was actually streamed and the endpoint signal sent. Network delay remains included.
- Context
- Latency is sample-weighted within datasets; Earnings22 chunks are aggregated back to parent files.
Artificial Analysis VoxPopuli-Cleaned-AA time to final transcript after endpoint (Show measurement, test conditions, and source)
- Source value
- 0.16497676305571973
- Score
- 0.16497676305571973
- Metric
- latency after detected end of speech
- Unit
- seconds
- Benchmark version
- AA-WER v2.2 streaming; retrieved 2026-10-07
- Category
- audio
- Direction
- Lower is better
- Sample
- 628
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis AA-WER Streaming
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. Exact Muse Voice Transcribe 1.0 row; English parliamentary speech Timer starts after the speech-end audio was actually streamed and the endpoint signal sent. Network delay remains included.
- Context
- Latency is sample-weighted within datasets; Earnings22 chunks are aggregated back to parent files.
Page 1 of 2
Published prices
- API price
- $0.18 per audio hour ($3 per 1,000 audio minutes)Source
Sources and data date
| Type | Evidence and data date |
|---|---|
| Research status | Research date October 7, 2026. 10 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Not published | diarization-sample-counts Original wording Meta's 17.5% average across AMI-IHM, AMI-SDM, and VoxConverse includes no per-dataset values, sample counts, or confidence intervals. |
| Not published | aa-language-coverage Original wording The AA streaming evaluation is English-language; the sources read publish no equivalent Muse WER/latency run for other languages. |
| Additional source | Meta AI Research Muse Voice Transcribe announcement (retrieved October 7, 2026) · Meta AI Research Muse Voice Transcribe announcement · Editorial description reviewed October 7, 2026 |
| Additional source | Meta Muse Voice Transcribe API and pricing (retrieved October 7, 2026) |
| Additional source | Artificial Analysis AA-WER Streaming (retrieved October 7, 2026) |
| Additional source | VoiceCodeBench released baseline (retrieved October 7, 2026) |
| Additional source | Model metadata (retrieved October 7, 2026) |
| Additional source | Model metadata (retrieved October 7, 2026) |
| Additional source | Model metadata (retrieved October 7, 2026) |
| Additional source | Model metadata (retrieved October 7, 2026) |
| Additional source | Model metadata (retrieved October 7, 2026) |
| Additional source | Model metadata (retrieved October 7, 2026) |