Stay connected with KayaToday, follow us on Instagram and Facebook for the latest news and reviews delivered straight to you.
The hardest problem in real-time transcription is not turning sound into words. Any reasonably modern model can do that. The hard problem is knowing who said which words, while the conversation is still happening, across a room full of people talking over each other. That is the problem Meta is now staking a commercial claim on.
Meta Superintelligence Labs launched Muse Voice Transcribe this week, a streaming audio model that bundles transcription, endpoint detection and speaker diarization for more than 20 simultaneous speakers into a single API priced at $0.18 per hour of processed audio, or $3 per 1,000 minutes. The model was trained across more than 70 languages, with 25 extensively validated at launch, and it supports long audio sessions exceeding one hour, multilingual code-switching mid-sentence, and language and keyword biasing.
Why Diarization Inside the Model Changes the Calculus
Speaker diarization has traditionally been a post-processing step bolted onto a transcription pipeline after the fact. A speech recognition engine produces a transcript, then a separate clustering algorithm tries to assign speaker labels to segments. The seam between those two steps is where errors compound, particularly in live settings where the system cannot look ahead at future audio to resolve ambiguity.
Muse takes a different approach by folding speaker attribution directly into its autoregressive architecture. Audio arrives in 80-millisecond chunks, which works out to 12.5 chunks per second, with each chunk transformed into a soft token. At each step the model decides whether to consume more audio or emit text. Meta calls this adaptive delay, meaning the model can hold back longer when speech is ambiguous and commit earlier when context is sufficient. Reinforcement learning combines word-error-rate and delay rewards to train that behavior, according to Meta’s technical explanation of the architecture.
Speaker turns are encoded in the same token sequence as the transcript itself. A <|start_of_turn|> token marks a potential new speaker turn, tokens such as <|speaker_A|> identify the participant, and separate onset and endpoint tokens mark speech boundaries. Because ASR, diarization and endpointing are trained together rather than chained as separate processes, the model does not need to reconcile outputs from systems that were optimized independently.
The practical consequence for enterprise developers is that speaker attribution does not require a separate pipeline to maintain or a separate billing line to absorb. Meeting intelligence tools, call-centre analytics platforms, compliance recording systems and ambient voice agents all depend on knowing who said what, not merely what was said. Getting that attribution wrong in a recorded board meeting or a regulated customer call is not a minor inconvenience.
The Speaker-Count Claim Deserves Scrutiny
Meta’s 20-plus-speaker figure sits toward the high end of the market, but it is not a record, and the distinction matters for anyone evaluating the product against alternatives.
Speechmatics currently makes the strongest explicit real-time capacity claim in the space. Its real-time documentation states that speaker diarization supports 50 speakers by default and can be configured up to 100. Amazon Transcribe specifies a maximum of 30 unique speakers for streaming transcription. Both exceed Meta’s stated ceiling.
At the other end, AssemblyAI’s streaming diarization caps at 10 speakers, and Soniox documents a maximum of 15 per session. Both companies note that live attribution is harder than offline processing because streaming systems must make decisions without access to future audio context, which is exactly the tradeoff Muse’s adaptive delay mechanism is designed to manage.
It is also worth noting that Meta’s own launch demonstrations do not stress-test the 20-plus figure. The principal live demo uses eight speakers, and its long-form recording features 11 labeled participants. The 20-plus number is a stated model capability, not the participant count shown in public material.
Speaker capacity and diarization accuracy are also separate questions. A platform that can represent 100 speakers is not automatically better at correctly attributing speech than one supporting 20. Meta’s benchmark data addresses accuracy directly: the company reports an average 17.5% diarization error rate across the AMI-IHM, AMI-SDM and VoxConverse evaluation sets, lower than the competing systems shown in its published chart. On the Artificial Analysis AA-WER Streaming Index supplied with the launch, Muse records a 3.1% final-transcription word error rate, ahead of Cartesia Ink-2 at 3.4%, ElevenLabs Scribe v2 Realtime at 3.6%, Qwen3 ASR Flash Realtime at 3.7%, GPT Live Transcribe and Grok Speech to Text Streaming at 3.9%, and Gemini 3.5 Transcribe Live and AssemblyAI U3.5 Realtime Pro at 4.0%. Artificial Analysis is an independent third-party benchmarking firm, and Meta says Muse took the number one spot on its streaming speech-to-text evaluation as of September 1.
Where $0.18 Per Hour Actually Sits in the Market
Pricing is where Muse’s competitive position becomes genuinely interesting. Standardising publicly posted rates to one hour of streaming audio with diarization included reveals a market with wide dispersion.
Soniox currently publishes a lower equivalent rate at roughly $0.12 per hour, though its diarization ceiling is 15 speakers. Meta at $0.18 undercuts xAI’s Speech to Text at approximately $0.20, Speechmatics Real-time Standard at around $0.24, and Qwen3 ASR Flash Realtime at roughly $0.324 for international deployments. Further up the scale, Deepgram Nova-3 Multilingual costs approximately $0.35 per hour for transcription alone, rising to around $0.47 once its $0.12-per-hour diarization add-on is included. AssemblyAI similarly lists $0.45 per hour for Universal-3.5 Pro Realtime and another $0.12 per hour for streaming diarization. OpenAI’s GPT Live Transcribe sits at $1.02 per hour and does not list diarization as a model capability.
The bundling point matters. Several competitors charge separately for speaker attribution, meaning a headline transcription rate understates the true cost for any application that actually needs to know who is speaking. Meta’s $0.18 covers both, and the company states that zero-data-retention processing is priced at parity with standard processing, which is relevant for regulated industries where data handling commitments affect vendor selection.
There are real limitations to acknowledge. The API currently provides turn-level timestamps rather than word-level timestamps, and it does not expose word-level confidence scores, sound-event detection or emotion detection. Sessions are capped at 60 minutes before an application must reconnect, and the default concurrent stream limit is eight per tenant. For developers building large-scale contact-centre infrastructure or multi-room meeting platforms, those constraints will require evaluation against operational requirements.
What This Means for Developers and the Competitive Landscape
For software teams in Malaysia and Singapore building voice-enabled products, the practical implication is a new reference point for what enterprise-grade real-time transcription should cost. The region’s growing base of SaaS companies, contact-centre operators and enterprise productivity tool builders all have use cases where speaker-attributed live transcription is either a core feature or a compliance requirement. At 1,000 hours of processed audio, Meta’s public rate implies roughly $180 in transcription charges, a figure that changes the unit economics of embedding voice intelligence into a product.
The broader competitive signal is that the speech-to-text market is moving toward bundled, low-latency, speaker-aware models rather than modular pipelines where accuracy, diarization and endpointing are purchased and integrated separately. Meta’s entry, backed by benchmark results that currently lead the Artificial Analysis streaming index, puts pressure on incumbents to compete on speaker-aware accuracy and total operating cost rather than raw transcription capability alone.
The 20-plus-speaker ceiling is not the most important number in Muse’s launch. The more important question is whether the model can maintain attribution, accurate text and usable turn boundaries while a complicated real-world conversation is still unfolding. Meta’s benchmark data suggests it can, and at $0.18 per hour, the cost of finding out is low enough that enterprise developers have little reason not to test it.
Read More: Your AI is not ignorant, it is forgetful. A Google study reframes the hallucination problem entirely
