Meta telah meluncurkan Muse Voice Transcribe, sebuah model pengenalan suara AI berbasis real-time yang mendukung lebih dari 70 bahasa, dapat menangani hingga 20 pembicara secara bersamaan, dan tersedia di Mac.
Meta, the tech giant, has unveiled Muse Voice Transcribe, a cutting‑edge AI voice recognition model built for real‑time transcription tasks.
Developers describe Muse Voice Transcribe as the first real‑time audio perception model introduced in Meta’s broader ecosystem.
The sophisticated software can identify over 20 speakers simultaneously and has been trained on more than 70 languages, 25 of which have undergone extensive verification.
In addition, Muse Voice Transcribe is built adaptively to handle multilingual conversations, instantly spotting language switches during dialogue and processing recordings that exceed one hour.
In various real‑time voice‑processing benchmarks, Muse Voice Transcribe reportedly outperforms competing models. According to Artificial Analysis’ AA‑WER Streaming speech‑to‑text benchmark, the system recorded a word error rate (WER) of 3.1 % for English conversations.
That figure is considerably lower than that of similar rivals such as Cartesia Ink‑2 (3.4 %), ElevenLabs Scribe v2 Real‑time (3.6 %), GPT Live Transcribe (3.9 %), and Gemini 3.5 Transcribe Live (4 %).
Cara Kerja Muse Voice Transcribe
Architecturally, Muse Voice Transcribe is a multimodal autoregressive model derived from the Muse Spark family. Its processing workflow divides the audio stream into 80‑millisecond segments (12.5 segments per second), each of which is compressed into a single soft token.
The smart system decides on its own whether to turn an audio chunk straight into a text token or to hold it temporarily while waiting for more context from the next segment.
The novel technique called adaptive delay is learned by the model using reinforcement learning methods.
Meta embeds two primary evaluation metrics—transcription accuracy and low latency—so the system is incentivized to generate text quickly without sacrificing meaning precision.
A comparable mechanism is also employed to enhance the accuracy of speaker separation (diarization).
Ketersediaan di Mac
Apple ecosystem users can now access Muse Voice Transcribe via the Meta Model API, the Meta AI macOS application, and Muse Code.
On a Mac, users simply hold down the Fn key to speak text directly into apps, or issue interactive voice commands through the Muse Code interface. Spencer Barnett also demonstrated the feature on his official X account.
For commercial application development, Meta charges an API usage fee of US$3 (approximately Rp 53,310) per 1,000 audio minutes, or about US$0.18 (around Rp 3,199) per hour.
Unlike the earlier Muse Glimmer model, Meta has chosen not to publish open‑weight versions of Muse Voice Transcribe, according to a report by The New Stack.
The arrival of this advanced transcription model is expected to intensify competition within the AI industry, especially in high‑level voice recognition.
To stay updated on the latest Muse Voice Transcribe features, cross‑platform device developments, and in‑depth global tech coverage, keep exploring more curated articles at TheCuy Tekno.
FAQ
1. Apa itu Muse Voice Transcribe?
Muse Voice Transcribe is Meta’s first real‑time AI voice recognition model, trained on over 70 languages and capable of identifying more than 20 speakers simultaneously.
2. Bagaimana sistem menentukan kecepatan transkripsi teks?
The system uses an adaptive delay approach based on reinforcement learning to assess whether an audio segment should be temporarily held to capture a fuller sentence context.
3. Platform apa saja yang mendukung penggunaan model ini?
The service can be accessed through the Meta Model API, the Muse Code development platform, and is directly integrated into the Meta AI app for macOS.
4. Berapa biaya akses API untuk model pengenalan suara ini?
Meta charges a fee of US$3 (approximately Rp 53,310) per 1,000 audio minutes, or about US$0.18 per hour, for API usage.
Terima kasih sudah mengunjungi thecuy. 🙂
Penulis Berpengalaman 5 tahun.