Now live New models for transcription and text-to-speech. AssemblyAI-STT-3.7 AssemblyAI-TTS-1
Insights & Use Cases

How accurate is speech-to-text in 2026?

How accurate speech-to-text is in 2026: WER vs entity error vs cpWER, dated AssemblyAI-STT-3.7 benchmarks against the field, and how to improve accuracy.

Abstract green mobius illustration

Written by

Kelsey Foster

Published on

18 November 2026

The best speech-to-text models reach about 95–98% word accuracy on clean audio in 2026, and roughly 92–96% on real calls. On AssemblyAI's November 2026 benchmarks, AssemblyAI-STT-3.7 posts 7.44% word error rate on accented call-center recordings, and AssemblyAI-STT-3.7-Realtime-Agent posts 4.53% word error rate on 12,460 voice-agent calls. Accuracy drops on accents, noise, overlapping speakers 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 November 17, 2026: pre-recorded = AssemblyAI-STT-3.7 (32 languages); streaming = AssemblyAI-STT-3.7-Realtime for live transcripts and AssemblyAI-STT-3.7-Realtime-Agent for voice agents (32 languages each). Universal-3.5 Pro, the previous pre-recorded model, was deprecated on November 17, 2026.

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.

Accuracy also covers more than the words. Modern systems 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:

Metric What it measures Best used for
Word error rate (WER) Share of words substituted, inserted or deleted General comparisons, human-readable transcripts
Semantic WER Whether meaning is preserved, judged by a reasoning model Voice agents and LLM pipelines
Entity error rate Accuracy on names, numbers, IDs, drug names Contact centers, medical, legal
cpWER Transcription plus speaker attribution Meetings and multi-speaker calls
Character error rate (CER) Character-level accuracy Languages without clear word boundaries, such as Japanese and Mandarin

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.

Test speech-to-text accuracy on your audio

Upload a file or stream your microphone and see how AssemblyAI-STT-3.7 handles real accents, noise and conversational speech. No code required.

Try playground

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

The most accurate model depends on what you're transcribing, so compare on the test closest to your audio. Every table names its source and date. Pre-recorded figures come from the AssemblyAI-STT-3.7 research post; streaming figures come from the AssemblyAI-STT-3.7 launch post (both November 2026). Full tables and methodology are on the AssemblyAI benchmarks page. If you're turning these numbers into a vendor shortlist, our guide to choosing a speech-to-text API has a weighted scorecard.

Pre-recorded: accented call-center audio

On AppTek's public call-center dialogues (526 recordings across 14 accents, scored with AppTek's own script, every system run in November 2026 at default settings), AssemblyAI-STT-3.7 posts 7.44% WER. That's ahead of Google Gemini 3.5 Transcribe (7.52%), Soniox (7.89%), Deepgram Nova-3 (8.59%), Speechmatics Enhanced (8.76%) and OpenAI GPT-4o Transcribe (10.82%). Universal-3.5 Pro scored 8.35% on the same set.

Pre-recorded: code-switching WER

Code switching (moving between languages mid-sentence, like Spanglish) is one of the harshest tests for a transcription model. Average WER over 16 datasets of English mixed with German, Spanish, French, Italian or Portuguese. Lower is better.

Model Average WER
AssemblyAI-STT-3.7 6.2%
ElevenLabs Scribe v3 6.8%
Soniox stt-async-v5 10.2%
Azure Batch Transcription 17.4%
Mistral Voxtral Mini Transcribe 2 17.4%
AWS Transcribe 22.9%
NVIDIA Parakeet TDT 0.6B v3 29.3%
Cohere Transcribe (language pinned) 29.4%
OpenAI GPT-4o Transcribe 44.7%
Deepgram Nova-3 49.4%
Gladia Solaria-1 50.4%

AssemblyAI-STT-3.7 and ElevenLabs Scribe v3 measured November 2026, others September 2026, each vendor's production API at default settings. Speechmatics, Gemini and Grok are left out because their APIs failed repeatedly on these datasets.

The spread is enormous. AssemblyAI-STT-3.7 has the lowest average WER of the systems measured, and the lowest on 12 of the 16 datasets, ahead of ElevenLabs Scribe v3, Soniox, Azure, AWS Transcribe, OpenAI, Deepgram and Gladia. Past the top two, every system is above 17%. Switching languages also tempts models to switch when the speaker didn't: on English benchmark audio, AssemblyAI-STT-3.7 cut characters from other writing systems from 161 to 24 per million against Universal-3.5 Pro.

Pre-recorded: speaker diarization

For meetings and multi-speaker calls, the model has to attribute words to the right person. Diarization error rate (DER) scores how much audio gets the wrong speaker; cpWER (concatenated minimum-permutation word error rate) scores transcription and speaker assignment together. Here's why cpWER is a better diarization metric than DER. Average DER over six English conversation datasets, each vendor's production pre-recorded API at default settings. Lower is better.

Model Average DER
AssemblyAI-STT-3.7 10.7%
ElevenLabs Scribe v3 12.0%
Speechmatics Enhanced 16.2%
Gladia Solaria-1 16.7%
AWS Transcribe 16.9%
Deepgram Nova-3 17.4%
Soniox stt-async-v5 24.2%

AssemblyAI-STT-3.7 and ElevenLabs Scribe v3 measured November 2026; other providers October 2026.

AssemblyAI-STT-3.7 has the lowest diarization error rate of the pre-recorded APIs measured, ahead of ElevenLabs Scribe v3, Speechmatics, Gladia, AWS Transcribe, Deepgram and Soniox. On cpWER, it fell from 29.3% to 23.5% on full English recordings against Universal-3.5 Pro with its previous diarization model. Short audio is where diarization models usually struggle, because there's little speech to learn each voice from: on 60-second clips, cpWER fell from 34.0% to 18.6%. Outside English, on conversational recordings in nine languages, average cpWER fell from 23.5% to 22.1% against Universal-3.5 Pro.

Streaming: WER on voice-agent calls

For voice agents, read speech is the wrong test. AssemblyAI measures streaming models on 12,460 English voice-agent calls from 522 speakers, recorded as telephony audio across three noise conditions, with every provider run through the same harness and no key terms or context for anyone (AssemblyAI-STT-3.7-Realtime-Agent research post, runs August to November 2026). Of the 26 streaming models we ran, AssemblyAI-STT-3.7-Realtime-Agent has the lowest word error rate and the lowest entity error rate, ahead of Soniox, ElevenLabs, Speechmatics, Google and Deepgram. Lower is better.

Model Word error rate Entity error rate
AssemblyAI-STT-3.7-Realtime-Agent 4.53% 13.2%
Soniox Realtime v6 6.02% 15.2%
ElevenLabs Scribe v3 Live 7.38% 16.9%
Meta Muse Voice Transcribe 7.48% 21.2%
Speechmatics Enhanced Realtime 8.13% 19.8%
Google Gemini 3.5 Transcribe Live 8.48% 19.2%
Deepgram Nova-3 8.64% 25.1%
Deepgram Flux EN 13.50% 28.8%

AWS Transcribe Streaming and Gladia Solaria aren't scored on this set because too many calls failed to return a transcript.

Entity errors matter most in practice: when a caller reads out an amount, the gap between getting it right and asking them to repeat it is the gap between a finished call and a frustrated one. On each entity class shown in the research post (phone numbers, email addresses, names, dates and times, numbers and currency), the Agent model makes fewer errors than every other provider. Against its predecessor, Universal-3.6 Pro Realtime, the Agent model cut number errors on Voice Code Bench by 80% (11.28% to 2.27%) and short replies transcribed wrong by 29%.

Streaming: live transcription on short-form English

For notetakers and captions, AssemblyAI-STT-3.7-Realtime is the model to compare. Mean WER over Common Voice, LibriSpeech clean and LibriSpeech other, each vendor's realtime API at default settings:

Model Short-form English WER
AssemblyAI-STT-3.7-Realtime 3.23%
AWS Transcribe Streaming 3.93%
ElevenLabs Scribe v3 Live 4.18%
Speechmatics Enhanced Realtime 4.69%
Google Gemini 3.5 Transcribe Live 5.04%
Soniox Realtime v6 6.52%
Deepgram Flux EN 6.53%
Deepgram Nova-3 7.46%

Read this table against the voice-agent table above: AWS Transcribe Streaming does well on read speech but isn't scored on the voice-agent calls, and Soniox Realtime v6 does better on the voice-agent calls than on read speech. One benchmark never tells the whole story.

See the numbers on your own audio

Run AssemblyAI-STT-3.7 or one of the realtime models on a file that represents your real use case. Free to start.

Sign up free

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 have specific controls for those conditions.

Streaming (AssemblyAI-STT-3.7-Realtime and -Realtime-Agent):

  • Conversation context: the model carries prior finalized turns forward automatically. On voice agents, pass what your agent just said with agent_context, so a one-word reply like "seven" gets read as an answer to "How many guests?"
  • Prompting and key terms: describe the call with prompt and list names or product terms with keyterms_prompt, and update both mid-session. On 20,000 real voice-agent calls, a detailed prompt cut WER by 21% and entity errors on names by 49% against no prompt, per the streaming prompting docs.
  • Voice Focus (Agent model only): voice_focus set to near-field for headsets and handsets or far-field for rooms, kiosks and drive-thrus. Even before you set it, the Agent model cut words added while the caller was silent by 47% against Universal-3.6 Pro Realtime.

Pre-recorded (AssemblyAI-STT-3.7):

  • Contextual prompting and key terms: prime the model with the domain, participant names or product vocabulary.
  • Native code switching across 32 languages: no configuration. 14 languages are new in this release, and some are already among its most accurate: Galician at 4.1% WER, Russian at 4.6% and Catalan at 5.1%.
  • Transcript style: disfluencies keeps or removes fillers and false starts, and format_text returns written-form or spoken-form numbers. Both default to true. If your reference transcripts spell numbers out, set format_text to false before you score WER, or you'll count formatting differences as errors.

The practical takeaway hasn't changed: always test with audio that represents your actual use case.

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.

WER also penalizes formatting choices. "I cannot" and "I can't" score differently even when both are correct, and "$12.50" scores as wrong against a reference that says "twelve fifty." 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.

Scenario Typical accuracy range Main challenges
Clean studio recording 95–98% Minimal background noise, clear speech
Video conference calls 85–92% Network compression, microphone quality
Phone conversations 80–88% Audio compression, line quality
Noisy environments 70–85% Background noise, multiple speakers
Heavily accented speech 75–90% Varies by accent and training data
Domain-specific content 80–95% Technical terminology, proper nouns

These ranges are editorial estimates across providers, not benchmark results. Leading models now beat the low end of most rows; 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.

Attribute Human transcription AI transcription
Accuracy (optimal conditions) Near-perfect Near-human with leading models
Speed Slower than real time Real time or faster
Scale Limited by workforce Hundreds of parallel jobs by default
Cost Higher per hour Well under $1 per hour for leading APIs
Contextual understanding Excellent Improving fast with context and prompting
Consistency Varies by transcriptionist Highly consistent

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 AssemblyAI-STT-3.7-Realtime-Agent from the Realtime Speech-to-Text API with your own LLM and AssemblyAI's Text-to-Speech API for voice output, orchestrated with LiveKit, Pipecat or your own loop.

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. That's why accented test sets like AppTek's exist.
  • 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, as the 6% to 50% spread in the table above shows.

Accuracy requirements by industry

Different use cases tolerate different error rates, and the right metric matters as much as the target. These targets are editorial guidance, not benchmark results.

Use case Target accuracy Primary metric
Voice agents 90%+ general, 95%+ for critical commands Semantic WER, entity accuracy, dropped short replies
Contact centers 90%+ automated, 85%+ agent assist WER plus entity accuracy on account numbers
Meeting transcription 88%+ readable, 92%+ searchable Normalized WER plus cpWER
Medical and legal 98%+ Missed entity rate on critical terms

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 is one parameter, domain: "medical-v1", on AssemblyAI-STT-3.7 and both 3.7 realtime models, covering medication names in English, Spanish, German and French. Measured on the previous flagship, it cut missed medical entities by about 20% against the base model. 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

  • Keyterms prompting: supply product names, acronyms and proper nouns. AssemblyAI-STT-3.7 accepts up to 1,000 words or phrases.
  • Contextual prompting: describe the audio's domain and what you already know about it, with no retraining required.
  • Agent context (streaming): pass the agent's last reply so the model reads short answers in context.
  • Audio hygiene: use lossless or lightly compressed formats and keep the sample rate at 16 kHz or higher.
  • Match your scoring to your output style: set format_text and disfluencies to match your reference transcripts before you compare WER.
  • 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. Rerun your test set whenever you change models; AssemblyAI accounts that don't set speech_models move from Universal-3.5 Pro to AssemblyAI-STT-3.7 automatically on December 8, 2026, so pin the model if you need to control the timing.

The future of speech-to-text accuracy

The next gains won't come from squeezing another point off LibriSpeech. They'll come from models that use what your application already knows: the question the agent just asked, the caller's account, the names on the meeting invite. The docs' prompting benchmark shows the size of that lever, with name errors cut nearly in half by a detailed prompt.

Evaluation is moving the same way. As transcripts flow straight into LLMs, teams score entities, speaker attribution and whether the downstream task succeeded, alongside WER. Those are the numbers worth tracking on your own audio.

Test accuracy with your own audio

Sign up free and run AssemblyAI-STT-3.7, AssemblyAI-STT-3.7-Realtime or AssemblyAI-STT-3.7-Realtime-Agent on the files that matter to you.

Sign up free

Frequently Asked Questions (FAQs)

How accurate is AssemblyAI speech-to-text?

On AssemblyAI's November 2026 benchmarks, AssemblyAI-STT-3.7 posts 7.44% WER on AppTek's accented call-center set and 6.2% average WER on code-switched audio. For streaming, AssemblyAI-STT-3.7-Realtime-Agent posts 4.53% word error rate and a 13.2% entity error rate on 12,460 voice-agent calls. Your accuracy depends on your audio, so test a representative file.

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

It depends on the audio. On AssemblyAI's November 2026 tests, AssemblyAI-STT-3.7 has the lowest code-switching WER of the systems measured (6.2%, ahead of ElevenLabs Scribe v3 at 6.8%), and on AppTek's accented call-center set it's ahead of Google, Soniox, Deepgram, Speechmatics and OpenAI. For voice agents, AssemblyAI-STT-3.7-Realtime-Agent has the lowest WER and entity error rate of the 26 streaming models we ran. 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. Judge WER on your own audio, since noise, accents and formatting choices can double it.

Is AI transcription accurate?

Yes, on clean and moderately noisy audio leading AI models reach about 92–98% word accuracy. Accuracy falls on heavy crosstalk, very poor audio and rare names, so many teams add key terms, contextual prompts and human review of low-confidence words for high-stakes work.

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

Real-time speech-to-text is now close to pre-recorded accuracy. AssemblyAI-STT-3.7-Realtime posts 3.23% WER on short-form English, and AssemblyAI-STT-3.7-Realtime-Agent posts 4.53% word error rate on 12,460 voice-agent calls. Passing a detailed prompt cut streaming WER by a further 21% on 20,000 voice-agent calls, per AssemblyAI's docs.

How accurate is multilingual or code-switched transcription?

Code-switched transcription varies widely by model. Across 16 datasets mixing English with five European languages, AssemblyAI-STT-3.7 averaged 6.2% WER and ElevenLabs Scribe v3 6.8%, while OpenAI GPT-4o Transcribe scored 44.7% and Deepgram Nova-3 49.4%. AssemblyAI-STT-3.7 handles 32 languages with native code switching and no configuration.