How accurate is speech-to-text in 2026?

The best speech-to-text models reach about 95–98% word accuracy on clean audio in 2026. On AssemblyAI's published benchmarks, Universal-3.5 Pro averages 4.35% word error rate on pre-recorded audio and Universal-3.6 Pro Realtime posts 5.19% on real voice-agent calls. Accuracy drops on accents, noise, and code-switching, so test on your own audio.

That headline number hides most of the real story. Accuracy swings sharply once you move to accented speech, background noise, language switching, and the messy conversational audio that voice agents hear. This guide covers where accuracy stands, how the leading models compare head to head, what moves the number, and how to improve it in production.

Current AssemblyAI models as of October 2026: pre-recorded = Universal-3.5 Pro (18 languages, released July 7, 2026); streaming = Universal-3.6 Pro Realtime (32 languages, the default since September 29, 2026). Both handle mid-sentence code-switching.

What is speech-to-text accuracy?

Speech-to-text accuracy is how precisely a model converts spoken words into written text, expressed as a percentage where 100% means an error-free transcript. It's the foundational metric for any speech-to-text system, and it decides whether your output is useful or frustrating.

The gap between scores matters more than it looks. An 85% accurate system makes about 15 errors per 100 words, so transcripts need heavy cleanup. A 95% accurate system makes about 5, often minor punctuation or formatting slips that don't block understanding.

But accuracy isn't only about the words. Modern systems also have to handle punctuation, casing, speaker changes, and context. A model can hear "there," "their," and "they're" correctly and still pick the wrong spelling.

How is speech-to-text accuracy measured?

The standard metric is word error rate (WER): substitutions, insertions, and deletions divided by the number of words in a human-verified reference transcript. Accuracy is roughly 100% minus WER. So a 5% WER means about 95% of words are right.

WER is useful but incomplete. It treats "um" the same as a phone number, and it inherits every mistake in the reference transcript. That's why teams pair it with other metrics:

For step-by-step instructions, normalization code, and how to automate WER checks, see our guide on how to measure transcription accuracy. We explain why WER alone misleads in word error rate is broken.

The most accurate speech-to-text models in 2026: a head-to-head

The most accurate model depends on what you're transcribing. On AssemblyAI's published benchmarks, Universal-3.5 Pro leads pre-recorded audio and Universal-3.6 Pro Realtime leads streaming. Here's how they compare with Deepgram, ElevenLabs, OpenAI, and others. Every table names its source and date.

Pre-recorded: average WER

Average WER across four dataset groups (synthetic medical, accented English from India, general speech, and webinar speech), from the AssemblyAI benchmarks page (accessed October 8, 2026). Lower is better.

Pre-recorded: code-switching normalized WER

Code-switching (switching languages mid-sentence, like Spanglish or Hinglish) is one of the harshest tests for a transcription model. This is normalized WER averaged across five language pairs, published with the Universal-3.5 Pro launch on July 7, 2026. Lower is better.

The spread is enormous. GPT-4o Transcribe's 44.58% is nearly six times Universal-3.5 Pro's error rate on the same task. Interestingly, GPT-4o Transcribe is competitive on the monolingual table above, which shows why one benchmark never tells the whole story.

Pre-recorded: speaker diarization (cpWER)

For meetings and multi-speaker calls, the model has to attribute words to the right person. cpWER (concatenated minimum-permutation word error rate) scores transcription and speaker assignment together. Here's why cpWER is a better diarization metric than DER. Averages across DiPCo, CALLHOME, NOTSOFAR, and AMI, from the benchmarks page (accessed October 8, 2026). Lower is better.

Universal-3.5 Pro leads, with Azure a close second. Most of the field sits 5–8 points behind.

Streaming: WER on voice-agent audio

For streaming and voice agents, AssemblyAI measures on an English voice-agent benchmark of 12,460 scripted caller scenarios across three noise environments, not read speech (published September 29, 2026). Lower is better.

Universal-3.6 Pro Realtime also improves on its predecessor: 5.19% WER versus 5.72% for Universal-3.5 Pro Realtime on the same set. Independent leaderboards point the same way. It posted a 0.96% pooled semantic WER on Pipecat's open STT benchmark (rendered September 28, 2026) and the lowest WER, 2.2%, on Coval's streaming leaderboard over a 7-day window at launch.

The entity rows are where the gap gets practical. When a caller reads out a phone number, the difference between a 2.4% and an 11.5% error rate is the difference between an agent that books the appointment and one that asks the caller to repeat themselves.

Customers notice the difference in their own products:

"On 10 out of 10 onboarding calls, our customers are at some point telling us 'wow that insight was crisp' — and that's because of the accuracy we're getting from AssemblyAI."

— Jake Cronin, Co-founder & CEO, Siro

Accuracy on accents, noise, and real-world audio

Clean benchmarks are the easy part. The accuracy you'll see depends on accents, background noise, crosstalk, and microphone quality. AssemblyAI's current models include specific features for those conditions.

Streaming features (Universal-3.6 Pro Realtime):

• Context carryover: a rolling memory of the conversation, on by default. Once a caller says an unusual name, the model is far less likely to mangle it the second time.
• Agent context: pass the agent's last question with agent_context so the model interprets the reply in context. Across 20,000 voice-agent files, this cut WER by 10.2%, with fabrications down 18.3%, hallucinations down 17.2%, and short-utterance errors down 13.7%.
• Voice focus: voice_focus isolates the primary speaker, with near_field for headsets and phones and far_field for rooms, kiosks, and drive-thrus.

Pre-recorded features (Universal-3.5 Pro):

• Contextual prompting: prime the model with the domain or prior context. In an internal healthcare test, feeding prior-visit notes cut missed medical terms by 31%. Metaview saw low-confidence tokens fall about 47% by supplying meeting metadata.
• Native code-switching: 18 languages, no configuration needed.

The practical takeaway hasn't changed: always test with audio that represents your actual use case. Benchmark scores are a starting point, not a verdict.

WER vs. real-world accuracy: reading benchmarks honestly

There's often a big gap between marketing numbers and production reality. Benchmarks use standardized audio; your users are on conference calls with spotty internet, in noisy cars, or on cheap mics. Teams building on Voice AI, like Veed, judge models by how they perform on their users' audio, not on a leaderboard.

WER also penalizes formatting choices. "I cannot" and "I can't" score differently even when both are correct. And human reference files contain inconsistencies that inflate WER for models that transcribe more faithfully. When you read a benchmark claim, ask how the reference was made and whether it was normalized. Those choices can shift WER by several points.

These ranges are editorial estimates across providers, not benchmark results. Your numbers will vary with your audio.

Human vs. AI accuracy: setting realistic expectations

Professional transcriptionists achieve near-perfect accuracy in good conditions. They bring a lifetime of context that helps them decipher mumbled words, crosstalk, and noise.

Modern Voice AI models have closed much of that gap, and they win on speed, scale, and consistency.

For real-time products, a human isn't an option at all. A voice agent needs the transcript now, and if it's wrong, the LLM responds to the wrong thing. You can pair Universal-3.6 Pro Realtime with your own LLM and text-to-speech, or use the Voice Agent API, which runs on Universal-3.6 Pro Realtime and bundles all three for $4.50/hr.

Factors that impact speech-to-text accuracy

Audio quality factors

• Microphone quality: built-in laptop mics usually score lower than headsets or USB mics.
• Background noise: traffic, air conditioning, and office chatter cause errors, especially for quieter speakers.
• Audio compression: low-bitrate MP3s and phone codecs introduce artifacts.
• Recording environment: hard surfaces add echo; soft furnishings improve clarity.

Speaker-related factors

• Accent and dialect: modern models handle variety far better than older ones, but gaps remain.
• Speaking pace: very fast or very slow speech reduces accuracy.
• Overlapping speech: crosstalk is hard for both transcription and speaker attribution.

Content and context factors

• Vocabulary: technical jargon scores lower than everyday language.
• Proper nouns: names of people, companies, and products are frequent failure points.
• Numbers and dates: "fifteen" vs. "fifty" and date formats need context.
• Language mixing: code-switching hurts most models; this is where Universal-3.5 Pro's native code-switching pulls ahead.

Accuracy requirements by industry

Different use cases tolerate different error rates, and the right metric matters as much as the target.

For contact centers, entity accuracy on spelled-out data like card and phone numbers drives conversion and compliance; see our contact center solutions. For medical transcription, AssemblyAI's Medical Mode cuts missed medical entities by about 20% versus the base model with one parameter. AssemblyAI is considered a business associate under HIPAA and offers a Business Associate Addendum (BAA) for customers processing PHI.

Improving speech-to-text accuracy in your applications

• Keyterm prompting: supply product names, acronyms, and proper nouns. Example: terms "AFib, GERD" turn "afib/gerd" into "AFib/GERD."
• Contextual prompting: describe the audio's domain and prior context, no retraining required.
• Agent context (streaming): pass the agent's question to cut WER on short replies.
• Audio hygiene: use lossless or lightly compressed formats and keep the sample rate at 16 kHz or higher.
• Confidence scores: flag words below a threshold (say 0.85) for human review on high-stakes workflows.

Once you're in production, treat accuracy as something you monitor, not something you check once. Our guide to measuring transcription accuracy shows how to set baselines and run automated WER regression tests in CI.

The future of speech-to-text accuracy

The most interesting shift isn't a smaller number on a clean dataset. It's that models are no longer passive listeners. Universal-3.6 Pro Realtime already reads the agent's question, carries context across turns, and isolates the speaker it's supposed to hear.

The next gains won't come from squeezing another point off LibriSpeech. They'll come from models that understand the situation they're transcribing. As transcripts flow straight into LLMs, evaluation is moving toward meaning preservation and task success. The accuracy that wins will be measured on your audio, in your conditions, doing your job.

Frequently asked questions

How accurate is AssemblyAI speech-to-text?

AssemblyAI's Universal-3.5 Pro averages 4.35% WER on pre-recorded audio across four dataset groups on AssemblyAI's published benchmarks, roughly 95–96% word accuracy. Its streaming model, Universal-3.6 Pro Realtime, posts 5.19% WER on a 12,460-scenario voice-agent benchmark and 0.96% pooled semantic WER on Pipecat's open benchmark. Your accuracy depends on your audio, so test a representative file.

What's the most accurate speech-to-text model in 2026?

On AssemblyAI's published benchmarks, Universal-3.5 Pro is the most accurate pre-recorded model measured (4.35% average WER, 7.69% code-switching WER, 30.17% cpWER). For streaming, Universal-3.6 Pro Realtime leads at 5.19% WER, ahead of ElevenLabs Scribe v2 (7.78%), Deepgram Nova-3 (8.64%), and Deepgram Flux EN (13.50%). No model wins every scenario, so benchmark on your own audio.

What is considered a good word error rate?

Under 5% WER is excellent and close to human transcription on clean audio. 5–10% is good for most production uses, and above 15% usually needs heavy correction. AssemblyAI's Universal-3.5 Pro averages 4.35% on its published benchmarks. Judge WER on your own audio, since noise and accents can double it.

How accurate is speech-to-text on accents and noisy audio?

Accuracy on accented and noisy audio typically ranges from about 70% to 90%, below the 95%+ you'll see on clean recordings. AssemblyAI's Universal-3.6 Pro Realtime narrows the gap for live audio with voice_focus and context carryover, posting 6.13% WER in heavy noise on its voice-agent benchmark. For pre-recorded audio, Universal-3.5 Pro uses contextual prompting to handle domain vocabulary.

How accurate is real-time speech-to-text?

Real-time speech-to-text is now close to batch accuracy. AssemblyAI's Universal-3.6 Pro Realtime posts 5.19% WER and a 14.4% entity error rate on 12,460 voice-agent scenarios. Passing the agent's question via agent_context cut WER a further 10.2% across 20,000 voice-agent files. Compare streaming models at the latency you'll actually run.

How accurate is multilingual or code-switched transcription?

Code-switched transcription varies widely by model. On AssemblyAI's five-language-pair benchmark, Universal-3.5 Pro averaged 7.69% normalized WER, while OpenAI GPT-4o Transcribe scored 44.58%. Universal-3.5 Pro handles native code-switching across 18 languages with no configuration, and Universal-3.6 Pro Realtime covers 32 languages for streaming.

Is AI transcription as accurate as human transcription?

On clean audio, leading AI models like AssemblyAI's Universal-3.5 Pro are close to professional human transcribers, and they're faster, cheaper, and more consistent. Humans still have an edge on heavy crosstalk, very poor audio, and context only a person in the room would know. Many teams use AI first and route low-confidence segments to human review.