Meta shipped a real-time transcription model on September 1, 2026, and quietly forced every other speech-to-text API vendor to defend their pricing. Muse Voice Transcribe, built by Meta Superintelligence Labs, landed at the top of the Artificial Analysis AA-WER Streaming leaderboard with a 3.1% word error rate and a price of roughly $0.18 per hour of audio. That is about a fifth of what Google Cloud charges for its standard speech-to-text tier.
This is a different fight from the one getting most of the headlines this month. OpenAI’s GPT-Live-1 and its full-duplex voice API are built for AI agents that talk back in real time, not for turning hours of audio into clean, searchable text. If you run a call center, build meeting software, or need accurate captions at scale, the speech-to-text API you pick matters more than which chatbot answers fastest. This piece compares the transcription-focused side of the market: Meta Muse Voice Transcribe, Google’s Chirp and Gemini 3.5 Transcribe Live, Deepgram Nova-3, and AssemblyAI’s Universal-3.5 Pro Realtime, on word error rate, latency, diarization, and cost per hour.
We pulled every number here from pricing pages, benchmark leaderboards, and launch coverage published in the first two weeks of September 2026. Where a figure was contested or unpublished, we say so rather than guess.
Speech-to-text sounds like a solved problem until you actually ship it. Audio quality varies wildly between a phone call, a laptop mic in a noisy office, and a professional headset. Every vendor claims strong accuracy in its marketing copy, and every vendor’s number comes from a test set chosen to make that vendor look good. That’s exactly why a neutral benchmark like AA-WER Streaming, cross-checked against real pricing pages, matters more than any single company’s press release.
What Changed in the Speech-to-Text API Market This Month
Speech-to-text has been a mature, almost boring corner of AI for years. Google, Amazon, and a handful of startups split the market, prices barely moved, and word error rate improvements came in fractions of a percent. That changed on September 1, 2026, when Meta Superintelligence Labs released Muse Voice Transcribe, described in its own materials as the lab’s first real-time audio perception model. Two days later, on September 2, it showed up inside Muse Code and Meta AI for Mac, meaning it went from a lab benchmark entry to a shipping product in under 72 hours.
The headline number is the AA-WER Streaming score of 3.1%, which put Muse ahead of Cartesia Ink-2 (around 3.4%) and ElevenLabs Scribe v2 Realtime (around 3.6%) on the same leaderboard. Google’s Gemini 3.5 Transcribe Live trailed at roughly 4.0%. On its own, a percentage point of word error rate sounds trivial. At call-center scale, where a single transcript feeds compliance checks, sentiment scoring, and CRM notes, a point of error is thousands of misread words a day.
Pricing moved just as fast. Muse Voice Transcribe launched at $3 per 1,000 audio minutes, or about $0.18 an hour. Meta’s own framing calls that roughly one-fifth of Google Cloud’s standard speech-to-text rate, and the math checks out: Google’s Chirp (Speech-to-Text V2) lists at $0.96 an hour, or $16 per 1,000 minutes. That is not a small discount. It is a five-fold gap between the newest entrant and the incumbent everyone assumed was the default choice for enterprise transcription.
None of this happened in isolation. The same week, OpenAI pushed GPT-6 Astra further into enterprise workflows, and Google had already dropped Gemini 3.8 Flash pricing to $0.75 per million tokens earlier in the month. Meta’s own voice push extends beyond transcription too, with custom voice generation quietly rolling into Muse alongside the transcription model. Speech and voice AI, across generation, transcription, and real-time conversation, has become the busiest corner of the model market this quarter.
Meet the Contenders
Four products anchor this comparison, plus two smaller players that show up on the same benchmark leaderboard. Each targets a slightly different slice of the speech-to-text API market, so the “best” answer depends heavily on your workload. A developer building a scrappy weekend project cares about very different things than a compliance officer signing off on a call-center vendor, and this section covers both ends of that spectrum.
Think of these four as sitting on two axes: price versus maturity, and accuracy versus ecosystem depth. Muse sits furthest toward price and accuracy, with the least track record. Google sits furthest toward maturity and ecosystem depth, at the highest price. Deepgram and AssemblyAI occupy the middle ground in different ways, one leaning on developer experience, the other on itemized pricing transparency. Keep that map in mind as you read the specs and pricing breakdowns below.
Meta Muse Voice Transcribe
Meta’s newest model handles streaming automatic speech recognition, speaker diarization, and endpointing (detecting when someone has actually stopped talking) in a single pass. It can separate more than 20 speakers in one audio stream and process multiple languages in the same session without a manual language switch. It ships through the Meta Model API, inside Meta AI for Mac as a dictation tool, and inside Muse Code for voice-driven code review and commit notes.
Google Chirp (Speech-to-Text V2) and Gemini 3.5 Transcribe Live
Google runs two products in this space. Chirp is the standard Google Cloud speech-to-text API most enterprise developers already integrate with, billed at $0.96 per hour. Gemini 3.5 Transcribe Live is the newer, benchmark-facing option, scoring closer to Muse on word error rate but still behind at roughly 4.0% on AA-WER Streaming. If you already run on Google Cloud infrastructure, the appeal is less about winning the benchmark and more about staying inside one billing account and one IAM setup.
Deepgram Nova-3
Deepgram has spent years building a developer-first reputation, with fast onboarding, generous docs, and a pricing page that is easy to actually read. Nova-3 streaming lists at $0.0077 per minute standard, or about $0.46 an hour, with a promotional rate around $0.0048 per minute (about $0.29 an hour) for qualifying volume, according to the Deepgram pricing page. Deepgram pricing sits in the middle of this comparison: cheaper than Google, more expensive than Muse, but with a mature SDK ecosystem that Meta’s newer API hasn’t had time to build yet.
AssemblyAI Universal-3.5 Pro Realtime
AssemblyAI’s realtime speech-to-text line bills transcription and diarization as separate line items. Universal-3.5 Pro Realtime transcription runs $0.15 an hour on its own. Add streaming diarization and the total climbs to about $0.27 an hour. That structure, visible on the AssemblyAI Universal Streaming product page, makes AssemblyAI transparent about what diarization actually costs, which matters if your use case is plain transcription without speaker labels.
Two more names round out the benchmark picture without becoming full comparison subjects here: Cartesia Ink-2 and ElevenLabs Scribe v2 Realtime, both close behind Muse on word error rate but with less public pricing detail available at the time of writing. Neither currently publishes a full self-serve pricing page the way Deepgram and AssemblyAI do, which makes them harder to budget for even where their accuracy numbers look competitive.
It’s worth sizing up the field this way because speech-to-text has quietly become a crowded category. A few years ago the realistic shortlist for any team building a transcription feature was Google, Amazon, and maybe one startup. Now there are at least six credible options with published benchmark scores, and picking wrong costs real money at scale, not just a worse demo.
Speech-to-Text API Specs Compared
The table below lines up the four main contenders across the specs that actually change an engineering decision: latency, speaker handling, language support, and access points.
| Attribute | Meta Muse Voice Transcribe | Google Chirp V2 / Gemini 3.5 Transcribe Live | Deepgram Nova-3 | AssemblyAI Universal-3.5 Pro Realtime |
|---|---|---|---|---|
| Developer | Meta Superintelligence Labs | Google Cloud / Google DeepMind | Deepgram | AssemblyAI |
| Launch date | September 1, 2026 | Chirp V2 established, Gemini 3.5 Transcribe Live benchmarked September 10, 2026 | Current mainline streaming model as of September 2026 | Current flagship streaming model as of September 2026 |
| AA-WER Streaming score | 3.1% | ~4.0% (Gemini 3.5 Transcribe Live) | Not published on AA-WER Streaming leaderboard | Not published on AA-WER Streaming leaderboard |
| End-of-utterance latency | 0.16 seconds after speech ends | Not separately published | Not separately published | Not separately published |
| Max simultaneous speakers identified | 20+ | Not specified in public materials | Speaker detection supported, cap not published | Diarization supported as add-on |
| Multilingual real-time support | Yes, multiple languages in one stream | Yes, broad language coverage via Google Cloud | Yes, multi-language streaming supported | Yes, multi-language streaming supported |
| Streaming support | Yes, native | Yes, native | Yes, native | Yes, native |
| Diarization included by default | Yes, built into base model | Varies by tier | Available, priced within core streaming rate | Priced as separate add-on ($0.12/hour) |
| Standard price per hour | ~$0.18 | $0.96 (Chirp V2) | ~$0.46 standard, ~$0.29 promotional | $0.45 (transcription only), $0.57 with diarization |
| Price per 1,000 minutes | $3 | $16 | ~$7.70 (standard rate, converted) | ~$7.50 (converted, transcription only) |
| Primary access points | Meta Model API, Meta AI for Mac, Muse Code | Google Cloud Console, Vertex AI | Deepgram Console, REST/WebSocket SDKs | AssemblyAI Console, streaming SDKs |
| Best-fit workload | Cost-sensitive, high-volume streaming with built-in diarization | Teams already standardized on Google Cloud | Developer-first integrations needing mature tooling | Workloads needing granular control over diarization billing |
A few of those rows stay blank on purpose. Deepgram and AssemblyAI have not published matching scores on the AA-WER Streaming leaderboard as of this writing, and inventing a number to fill the gap would defeat the point of a specs table. Where public data doesn’t exist yet, we say so.
Word Error Rate and Latency Benchmarks
Word error rate is the standard yardstick for any speech to text API, and the AA-WER Streaming benchmark run by Artificial Analysis is currently the closest thing the industry has to a neutral, cross-vendor leaderboard for real-time transcription. It measures how accurately a model transcribes live audio, not pre-recorded files, which matters because streaming conditions (background noise, overlapping speakers, network jitter) tend to push error rates higher than batch transcription benchmarks suggest.
| Model | AA-WER Streaming Score | Reported Latency | Source |
|---|---|---|---|
| Meta Muse Voice Transcribe | 3.1% | 0.16s after speech ends | Artificial Analysis, via MarkTechPost launch coverage |
| Cartesia Ink-2 | ~3.4% | Not separately published | Artificial Analysis leaderboard |
| ElevenLabs Scribe v2 Realtime | ~3.6% | Not separately published | Artificial Analysis leaderboard |
| Google Gemini 3.5 Transcribe Live | ~4.0% | Not separately published | Artificial Analysis leaderboard |
Three sources back this data: the Artificial Analysis benchmark itself, the launch write-up from MarkTechPost, and pricing comparisons published by industry trackers in the September 10, 2026 window. None of these numbers come from vendor marketing pages alone, which matters because self-reported benchmarks tend to flatter whoever ran the test.
Deepgram and AssemblyAI aren’t absent from this table because they perform badly. They simply haven’t published comparable scores on this specific leaderboard, and their own marketing pages use different test sets that don’t convert cleanly into an AA-WER Streaming number. If word error rate on this exact benchmark is your deciding factor, Muse currently wins. If your workload cares more about SDK maturity or ecosystem support, that single number matters less.
Speech-to-Text API Pricing Compared
Pricing across these four providers isn’t structured the same way, which makes a like-for-like comparison harder than it should be. Some bill per minute, some per hour, some bundle diarization and some charge for it separately. We converted everything into a common per-hour and per-1,000-minute rate below.
| Provider / Plan | Price per Minute | Price per Hour | Price per 1,000 Minutes |
|---|---|---|---|
| Meta Muse Voice Transcribe | $0.003 | ~$0.18 | $3 |
| Deepgram Nova-3 (promotional) | $0.0048 | ~$0.29 | $4.80 |
| AssemblyAI Universal-3.5 Pro Realtime (transcription only) | ~$0.0025 | $0.15 | ~$2.50 |
| Deepgram Nova-3 (standard) | $0.0077 | ~$0.46 | $7.70 |
| AssemblyAI Universal-3.5 Pro Realtime (with diarization) | ~$0.0045 | $0.27 | ~$4.50 |
| Google Chirp (Speech-to-Text V2) | $0.016 | $0.96 | $16 |
Run the math on a real workload and the gap stops being abstract. A contact center transcribing 50,000 minutes a month pays $150 with Muse, roughly $240 with Deepgram’s promotional rate, about $225 with AssemblyAI including diarization, and $800 with Google’s standard Chirp pricing. That is a $650-a-month swing for the exact same volume of audio, before anyone even accounts for accuracy differences.
Deepgram pricing sits in an interesting spot here. The promotional Nova-3 rate undercuts AssemblyAI’s diarization-inclusive tier, but only if you qualify for the discount, which typically depends on committed monthly volume. Standard Deepgram pricing without a volume commitment lands closer to AssemblyAI’s transcription-only rate. Read the fine print on any “promotional” number before building it into a budget.
Diarization and Speaker Handling
Diarization, the process of figuring out who said what in a multi-speaker recording, is where these four products differ the most in philosophy. Muse Voice Transcribe treats it as a core feature, not an upsell, and Meta’s own materials claim it separates more than 20 distinct speakers within a single real-time stream. For newsroom-style panel recordings, large conference calls, or classroom transcription, that speaker cap matters more than a fraction of a percent on word error rate.
AssemblyAI takes the opposite approach and prices diarization as a distinct line item ($0.12 an hour on top of transcription). That’s not necessarily worse. It gives finance teams a clean way to see exactly what speaker separation costs, and it lets teams that only need plain transcription skip the feature entirely and save money. Deepgram folds speaker detection into its core streaming rate without a separate charge, though the exact concurrent-speaker ceiling isn’t published in the same detail Meta provides.
Google’s documentation covers diarization across its speech products but doesn’t break out a distinct price for it within Chirp V2, and coverage varies by which specific API tier you provision. If your workload depends heavily on accurate speaker labels (legal depositions, multi-party interviews, healthcare intake calls) test each vendor against your actual audio before committing. Diarization accuracy on marketing pages rarely matches real-world audio with crosstalk and accents.
Why Word Error Rate Alone Doesn’t Tell the Whole Story
It’s tempting to treat the AA-WER Streaming leaderboard as the final word on which speech-to-text API to pick. Resist that urge. Word error rate is measured against a specific test set, recorded under specific conditions, and every vendor’s real-world traffic looks nothing like a benchmark clip. A model that scores 3.1% on clean, single-speaker English audio can slip several points higher on a crackly conference-room recording with three people talking over each other and a dog barking in the background.
Domain vocabulary matters just as much. Medical dictation, legal depositions, and software engineering standups all use words and acronyms that general-purpose speech models rarely see in training data. A generic 3.1% WER score says nothing about whether a model correctly transcribes “HIPAA,” “kubectl,” or a patient’s prescribed dosage. Several of these providers, including Deepgram and AssemblyAI, offer custom vocabulary and model fine-tuning specifically because the generic benchmark number undersells how much accuracy varies by industry.
Accent coverage is another blind spot. Benchmark leaderboards skew toward the languages and accents best represented in training data, typically US and UK English first. Teams with global user bases (call centers in the Philippines, support lines in Nigeria, meetings with non-native English speakers) should run their own accuracy tests before trusting a single published percentage. The gap between a benchmark score and real production accuracy has sunk more than one migration that looked great on paper.
Enterprise Considerations: Compliance and Data Handling
Price and word error rate get the headlines, but for regulated industries, compliance posture often decides the shortlist before accuracy even enters the conversation. Google Cloud carries the deepest bench of compliance certifications among the four, an advantage of a decade-plus of enterprise cloud infrastructure. Teams already running HIPAA-covered workloads or handling EU customer data frequently default to Google or a similarly established provider precisely because the paperwork is already done.
Deepgram and AssemblyAI, as more mature API-first companies, have each built out SOC 2 and similar attestations over the past several years, which is part of why they remain the safer middle-ground choice for startups that need enterprise credibility without a full Google Cloud contract. Meta’s Muse Voice Transcribe, being weeks old, simply hasn’t had time to accumulate the same audit history yet. That doesn’t mean it’s unsafe. It means procurement and legal teams at larger companies will likely ask for documentation that doesn’t exist yet, which can stall an otherwise attractive migration.
Data residency is the other question worth asking early. If your compliance requirements mandate that audio never leaves a specific region, confirm each vendor’s regional processing options before signing anything, since not every provider offers the same regional footprint for every tier of service. This is a case where the cheapest speech-to-text API on paper can turn into the most expensive one if it forces a redesign six months into a contract. Ask every vendor directly where audio is processed and stored, how long retention lasts by default, and whether you can force immediate deletion, since answers vary more than most pricing pages let on.
Real-World Examples
Numbers on a spec sheet only mean so much until you see where they actually get used. Here are five concrete scenarios where the choice of speech-to-text API changes the outcome.
- Voice-driven code review. Meta built Muse Voice Transcribe directly into Muse Code, its agentic coding environment, letting developers dictate commit messages and review comments instead of typing them, with the low-latency endpointing meaning the tool doesn’t cut off mid-sentence.
- Desktop dictation. Meta AI for Mac uses the same underlying model for real-time dictation, competing directly with Apple’s and Microsoft’s built-in dictation tools on accuracy rather than convenience alone.
- Contact center transcription at scale. A support operation handling tens of thousands of call minutes a month feels a five-fold price gap between Muse and Google Chirp far more than a one-point WER difference, since volume amplifies cost far faster than it amplifies error count.
- Multi-speaker meeting and panel transcription. Any product transcribing conference calls, webinars, or recorded panels needs reliable diarization across a dozen or more speakers, which is exactly the scenario Muse’s 20-plus speaker separation targets.
- Multilingual customer support. Teams fielding calls in more than one language in the same queue benefit from models that switch languages mid-stream without a manual toggle, a capability all four vendors now support to varying degrees.
Use-Case Recommendations
No single speech-to-text API wins every workload. Here’s how to match the product to the job. Think about your actual traffic pattern first: a bursty startup with unpredictable volume has different needs than an enterprise call center running near-constant load, and the cheapest per-minute rate doesn’t always translate into the cheapest total bill once you factor in engineering time, support response, and how often you’ll need to escalate an edge case.
- High-volume, cost-sensitive transcription: Meta Muse Voice Transcribe. At $0.18 an hour with diarization included, it’s hard to beat on pure unit economics, especially for startups processing large call volumes on tight margins.
- Teams already standardized on Google Cloud: Chirp V2 or Gemini 3.5 Transcribe Live. The accuracy gap is real but not disqualifying, and staying inside one cloud provider’s billing, IAM, and compliance certifications often outweighs a few cents per minute.
- Developer-first products needing mature SDKs and fast support: Deepgram Nova-3. Deepgram’s documentation and community tooling are more battle-tested than Meta’s newer API, which matters for small teams without dedicated infra engineers.
- Workloads needing granular diarization billing control: AssemblyAI Universal-3.5 Pro Realtime. Separating transcription and diarization costs makes budgeting predictable when speaker labels are optional per project.
- Regulated industries needing an established compliance track record: Google Cloud or Deepgram, both of which have years of enterprise audit history that a model launched two weeks ago simply hasn’t accumulated yet.
- Real-time captioning for live events: Whichever provider hits sub-second latency in your own testing. Muse’s published 0.16-second endpointing is the fastest publicly documented number here, but test it against your actual network conditions before betting a live broadcast on it.
How to Migrate Between Speech-to-Text APIs
Switching speech-to-text providers is less painful than migrating an LLM-based product, mostly because the input (audio) and output (text plus timestamps) are fairly standardized. Still, a rushed migration breaks diarization formatting, timestamp precision, and language-detection defaults. Teams that skip a proper parallel test tend to discover these gaps in production, usually during a customer escalation, which is the worst possible time to find out your new vendor formats speaker labels differently. Here’s a practical sequence.
- Audit your current usage. Pull three months of billing data from your existing provider and break it down by minutes transcribed, languages used, and how often diarization actually gets requested.
- Run a parallel accuracy test. Feed the same 30-60 minutes of real (anonymized) audio through both your current provider and the candidate API. Compare word error rate on your actual content, not the vendor’s demo clips.
- Check latency under your real network conditions. A model with a published 0.16-second endpointing figure in a lab test may behave differently over a mobile network or through a VPN. Test from the environment your users actually connect from.
- Map diarization output formats. Speaker labels, timestamps, and confidence scores rarely match exactly between vendors. Build a normalization layer before you route production traffic through the new API.
- Run a shadow deployment. Send a percentage of live traffic to the new provider while keeping your existing one as the source of truth, then compare outputs for at least two weeks before cutting over fully.
- Renegotiate volume pricing before switching, not after. Every vendor here has some form of committed-volume discount. Get a quote in writing before your existing contract renews so you have leverage either way.
- Cut over gradually by workload type, starting with the lowest-risk use case (internal meeting notes, for example) before moving customer-facing transcription over.
Pros and Cons
Meta Muse Voice Transcribe
Pros: Lowest published price per hour in this comparison, top score on the AA-WER Streaming leaderboard, diarization for 20-plus speakers included by default, fast endpointing at 0.16 seconds.
Cons: Brand new as of September 2026, meaning limited third-party SDK support, no long enterprise compliance track record yet, and documentation that’s still catching up to more established providers.
Google Chirp V2 / Gemini 3.5 Transcribe Live
Pros: Deep integration with the rest of Google Cloud, established enterprise compliance certifications, broad language coverage, predictable billing inside an existing GCP account.
Cons: Highest price per hour of the four, trailing word error rate on the current AA-WER Streaming benchmark, and a standard Chirp tier that hasn’t kept pace with newer streaming-native competitors on latency.
Deepgram Nova-3
Pros: Mature SDKs, developer-friendly documentation, mid-tier pricing with a meaningful promotional discount at volume, solid reputation among teams building voice products.
Cons: No published score on the AA-WER Streaming leaderboard for direct comparison, and promotional pricing depends on committed volume that smaller teams may not hit.
AssemblyAI Universal-3.5 Pro Realtime
Pros: Transparent, itemized pricing that separates transcription from diarization, giving finance teams precise cost control, plus a realtime-focused product built specifically for streaming use cases.
Cons: Diarization as an add-on pushes total cost above Deepgram’s promotional rate and well above Muse, and like Deepgram, it lacks a published AA-WER Streaming score for apples-to-apples comparison.
Sample Integration: Calling a Streaming Speech-to-Text API
Most of these providers follow a similar pattern for streaming transcription: open a WebSocket connection, stream raw audio chunks, and receive incremental transcript events back. Here’s a generic example of what that integration looks like in Python, using a WebSocket client against a streaming speech-to-text API endpoint.
import asyncio
import websockets
import json
async def transcribe_stream(audio_chunks, api_key, endpoint_url):
headers = {"Authorization": f"Bearer {api_key}"}
async with websockets.connect(endpoint_url, extra_headers=headers) as ws:
# Kick off the session with model + language config
await ws.send(json.dumps({
"type": "session.start",
"language": "auto",
"diarization": True,
"max_speakers": 20
}))
async def send_audio():
for chunk in audio_chunks:
await ws.send(chunk)
await asyncio.sleep(0.02)
await ws.send(json.dumps({"type": "session.end"}))
async def receive_transcripts():
async for message in ws:
event = json.loads(message)
if event.get("type") == "transcript.partial":
print("Partial:", event["text"])
elif event.get("type") == "transcript.final":
speaker = event.get("speaker_id", "unknown")
print(f"Speaker {speaker}: {event['text']}")
await asyncio.gather(send_audio(), receive_transcripts())
The exact message format differs across providers (Deepgram, AssemblyAI, and Meta’s Model API each use slightly different event names), but the shape of the integration stays consistent: stream audio in, receive partial and final transcript events back, and use the diarization flags to get speaker-labeled output. Budget engineering time for adapting this pattern regardless of which speech to text API you land on. If you’re piping transcripts into a larger retrieval system, the same streaming pattern shows up in most RAG pipeline builds, where transcript chunks get embedded and indexed as they arrive rather than after the full recording finishes.
The Verdict
On pure numbers, Meta Muse Voice Transcribe wins this comparison. It tops the AA-WER Streaming benchmark at 3.1%, undercuts Google’s Chirp pricing by roughly five times, and bundles diarization for more than 20 speakers without an extra line item. For any team building or migrating a high-volume transcription pipeline in September 2026, it’s the number to beat.
But “wins on the benchmark” and “right choice for your product” aren’t the same thing. Muse Voice Transcribe is two weeks old at the time of writing. Google Chirp, Deepgram Nova-3, and AssemblyAI’s Universal line have years of production hardening, documented compliance certifications, and SDKs battle-tested by teams that have already hit the edge cases you haven’t found yet. If your product needs to ship this quarter and can’t absorb a rocky integration with a brand-new API, Deepgram pricing (particularly the promotional Nova-3 rate) is the safer middle ground between cost and maturity.
Weigh these four factors before committing to any single provider: cost at your actual volume, accuracy on your actual audio (not a demo clip), how much diarization you genuinely need, and how much risk your team can absorb from a newer, less-proven API. A scrappy startup shipping a side feature can gamble on Muse’s price and accuracy lead. A healthcare company transcribing patient calls probably can’t, at least not yet, and should weight compliance history above the benchmark score.
The real story here isn’t which vendor wins today. It’s that Meta just proved a speech-to-text API can be both the most accurate and the cheapest option on the market at the same time, and that puts real pressure on Google, Deepgram, and AssemblyAI to move on price, accuracy, or both before their customers start running the same comparison you just did. For a broader view of how this fits into the rest of the model market, see our roundup of the best AI models of 2026.
Frequently Asked Questions
What is the most accurate speech-to-text API in 2026?
As of September 2026, Meta Muse Voice Transcribe holds the top score on the Artificial Analysis AA-WER Streaming leaderboard at 3.1% word error rate, ahead of Cartesia Ink-2 (~3.4%), ElevenLabs Scribe v2 Realtime (~3.6%), and Google’s Gemini 3.5 Transcribe Live (~4.0%). Deepgram and AssemblyAI have not published comparable scores on that specific leaderboard.
How much does the Google speech-to-text API cost?
Google’s Chirp (Speech-to-Text V2) is listed at $0.96 per hour of audio, or $16 per 1,000 minutes, based on pricing comparisons published in September 2026. That’s the highest per-hour rate among the four providers compared here.
Is Deepgram cheaper than AssemblyAI?
At promotional rates, yes. Deepgram Nova-3’s discounted rate runs about $0.29 an hour, versus AssemblyAI Universal-3.5 Pro Realtime at $0.45 an hour for transcription alone, or $0.57 an hour with diarization included. At standard (non-promotional) rates, Deepgram and AssemblyAI’s transcription-only pricing land much closer together.
Does Meta’s Muse Voice Transcribe support multiple languages?
Yes. Meta’s launch materials describe support for multiple languages within the same real-time stream, without requiring a manual language switch, alongside speaker diarization for more than 20 distinct voices.
What’s the difference between a speech-to-text API and a conversational voice API like GPT-Live-1?
A speech-to-text API like Muse Voice Transcribe, Chirp, Nova-3, or Universal-3.5 Pro Realtime turns spoken audio into written text, often with speaker labels. A conversational voice API, such as OpenAI’s GPT-Live-1, is built for full-duplex, real-time back-and-forth conversation, where the model listens and speaks at the same time. They solve different problems: one produces transcripts, the other powers talking AI agents. For a deeper look at conversational voice products, see our comparison of ChatGPT Voice, Grok Voice, and Gemini Live.
Can I use these speech-to-text APIs for call center transcription?
Yes, and cost differences matter more at call-center scale than almost anywhere else. A contact center transcribing 50,000 minutes a month would pay roughly $150 with Muse Voice Transcribe versus around $800 with Google Chirp for the same volume, a gap worth running the numbers on before committing to a vendor.
How accurate is diarization on these models?
Meta publishes a specific claim of separating more than 20 speakers per stream for Muse Voice Transcribe. Google, Deepgram, and AssemblyAI all support diarization, but public documentation doesn’t break out an exact speaker-count ceiling for each in the same detail. Test diarization accuracy against your own audio (with real crosstalk and accents) before assuming vendor marketing numbers will hold up in production.
Should I migrate away from Google Chirp right now?
Not without testing first. Google’s pricing is the highest in this comparison, but its enterprise compliance history and deep Google Cloud integration carry real value for regulated industries. Run a parallel accuracy and latency test against your actual audio, as outlined in the migration guide above, before moving production traffic to any newer, cheaper API.
Do these speech-to-text APIs support custom vocabulary?
Deepgram and AssemblyAI both offer custom vocabulary and keyword-boosting features that let you improve recognition of industry-specific terms, product names, or acronyms that a general-purpose model tends to mishear. Google Cloud supports similar customization through its speech adaptation tools. Public documentation for Meta’s Muse Voice Transcribe does not yet detail an equivalent custom vocabulary feature, since the product only launched on September 1, 2026.
What happens if a speech-to-text API mishears sensitive information?
Every provider in this comparison handles transcription errors differently depending on your integration, but the underlying risk is the same. If a transcript feeds automated decisions (compliance flags, billing codes, medical notes) an unreviewed error can propagate downstream. Build a human-review step for any workflow where a mistranscribed number or name carries real consequences, regardless of which vendor’s word error rate looks best on a leaderboard.
