Automatic speech recognition
In one sentence
Automatic speech recognition (ASR), also called speech to text, is the technology that turns spoken audio into written text, and it is the first real step in almost every voice agent.
Not to be confused with Voice biometrics and anti-spoofing, or Speech-to-speech models.
Definition
Automatic speech recognition converts spoken audio into written text. Speech to text is the plainer name for the same thing.
It is the first real step in almost every voice agent, and everything downstream inherits whatever it gets wrong.
Modern speech recognition is neural and end to end. Where the older arrangement needed an acoustic model, a pronunciation dictionary and a language model bolted together, one sequence-to-sequence architecture now does the job, usually a transformer, taking audio features in and emitting text tokens straight out. That is why accuracy climbed steeply from around 2020 onward, and why the technology copes with accents, disfluency and a noisy room far better than whatever people were using when they made up their minds about it.
Streaming against batch, the decision that shapes the product
- Batch transcription receives the complete audio and returns a full transcript. It is more accurate, because the model can see the whole utterance, including the words that came after the ambiguous one.
- Streaming transcription emits partial results as audio arrives and revises earlier guesses as context accumulates. Less accurate, and the only workable option for conversation, because you cannot wait for someone to finish before you start thinking.
- There are two kinds of output, interim and final. The interim ones wobble and get rewritten as more audio arrives. Wiring product logic to them is a mistake plenty of teams make and every one of them regrets.
Where accuracy actually goes wrong in production
- Proper nouns. Product names, place names and personal names are exactly the words a business voice agent most needs, and exactly the ones a general model handles worst. Custom vocabulary and keyword boosting exist for this.
- Numbers, and strings of them in particular. Postcodes, order numbers, phone numbers, amounts of money.
- Accents that are underrepresented in the training data.
- Overlapping speech, background conversation, and cheap laptop microphones.
- Code-switching, meaning a language change mid-sentence. It is extraordinarily common in multilingual markets and handled poorly by single-language models.
Adjacent terms and where the boundary falls
- Speech recognition is the act. Transcription is the output artifact.
- Diarization separates who spoke. It is a different task that often runs alongside.
- Speaker identification ties a voice to a person already on file. Different task again, and much nearer to biometrics.
- Natural language understanding is what happens to the text afterwards, and is not part of ASR.
Common misconception
That accuracy is one figure belonging to a vendor. It belongs to a combination instead: which model, what the audio sounded like, which words the domain uses, and who was doing the talking. A provider out in front on clean benchmark English can fall a long way behind on a noisy phone line in an accent it has not heard much of. Take a published word error rate as the point an evaluation starts from.
Why it matters commercially
Recognition caps everything that happens after it. Mishear the question, and the system retrieves the wrong passage and delivers a grounded answer that is wrong and sounds certain, so in a voice RAG stack the transcription error and the retrieval error stack on top of one another. Buyers who interrogate the language model and wave the speech recognition through are tuning the wrong half.
In voice specifically
In a text chat the words arrive exactly as the visitor typed them, so there is no recognition step and no error floor underneath the retrieval. Speech puts one there. Every downstream number, including retrieval quality and answer accuracy, is capped by what the transcript got right.
Where AsqVox fits
The Orb is live in English across global accents. For a website widget, accent breadth counts for more than raw benchmark accuracy, because visitors to a website are not a curated test set. Hindi and Hinglish, and with them code-switching support, are on the roadmap and not live.
Visual
The same sentence, four conditions
The benchmark is measured in Row 1. Your customers live in Rows 2 to 4.
Statistics
Every figure carries its source and year. Vendor numbers are labelled as vendor numbers, and where no reliable figure exists this page says so rather than borrowing one.
Whisper Large-v3 records around 2.7 percent word error rate on the clean LibriSpeech test-clean benchmark, and roughly 8 to 12 percent on real-world English audio.
2.7% clean, 8 to 12% real worldindependentWhisper Large-v3 on LibriSpeech test-clean, against reported real-world English performance, 2026 - The benchmark half is a published result on a public test set. The real-world half is a reported band rather than a single measurement, so quote it as a range and never as a provider figure.
Deepgram Nova-3 reports around 5.26 to 6.84 percent word error rate under production conditions.
5.26% to 6.84%vendor claimDeepgram Nova-3 reported figures, 2026 - Published by the provider. It is a production-conditions figure rather than a clean-benchmark one, which makes it more useful and still not comparable to another vendor number unless the test set and normalization match.
Benchmark accuracy and production accuracy are separated by roughly three to four times. This is the single most decision-relevant fact about speech recognition, and it is routinely left out of vendor material.
3x to 4xindustry rangeDerived from the benchmark and production figures above, 2026 - Directional. It is the gap to plan around, and the question to ask a vendor who quotes a benchmark number without one.
Testing by Coval and Gradium in 2026 found providers trading speed off against accuracy, and none of them ahead on both measures at once.
No provider leads on bothindependentCoval and Gradium benchmark, 2026 - This contradicts the common vendor claim of being simultaneously fastest and most accurate. It also means any accuracy comparison that ignores latency is incomplete.
Deepgram raised USD 130m at a USD 1.3bn valuation in January 2026.
USD 1.3bnindependentDeepgram funding announcement, January 2026, 2026 - A signal of how much capital the category is attracting, not a statement about accuracy. Funding is a proxy for conviction, not for performance.
Nothing clean, independent and public measures recognition accuracy broken out by accent or dialect, and that is the figure any consumer-facing deployment needs most.
-no reliable figureState the absence rather than filling it. If a vendor offers an accent figure, ask for the test set, the number of speakers and who recorded it.
Accuracy benchmarks for Indian languages and for Hinglish code-switching are just as thin. Vendors make claims. Almost nobody independent has checked them.
-no reliable figureRelevant to any deployment in a multilingual market, where mid-sentence language switching is normal speech rather than an edge case.
Examples
In practice
A retailer's voice agent keeps mishearing its own flagship product name and transcribes it three different wrong ways. Retrieval misses every time and the agent answers about a different product. The fix is not a better model. It is loading the product catalog as custom vocabulary with keyword boosting, after which the term transcribes reliably and retrieval lands. An afternoon of engineering time. Immediate effect on revenue.
The everyday version
Speech recognition is the bit that takes down what the customer just said. Where your company name is unusual, or your products are called things that are not ordinary words, expect it to mangle them until you hand over the list. Give it the list and the guessing stops.
Usage
Who says it
- Engineers say ASR, and will want to know straight away whether it is streaming or batch, which provider, and whether custom vocabulary is supported.
- Product and business people say speech to text, or just transcription.
- Both are correct and interchangeable. ASR came out of academia first; STT is what the market settled on.
Where it turns up
- It shares a clause with the language and accent list, streaming latency, word error rate, whether custom vocabulary and boosting are available, punctuation and formatting, redaction of personal data, and how long audio is kept.
- Retention is the clause procurement fixes on, because in most jurisdictions raw audio counts as personal data.
Common misuse
- Quoting a benchmark word error rate as though it predicted production accuracy.
- Assuming recognition quality is a solved commodity. Nothing else in the stack depends so heavily on the conditions it runs in.
- Confusing speech recognition with natural language understanding. Getting the words down is not understanding the request.
Questions people ask
What is the difference between ASR and speech to text?
Nothing. They name the same technology. ASR, automatic speech recognition, is the older academic term and speech to text, or STT, is the commercial one. Engineers tend to say ASR; product and business people tend to say speech to text or transcription.
How accurate is speech recognition?
It depends on conditions more than on the provider. Whisper Large-v3 records around 2.7 percent word error rate on the clean LibriSpeech benchmark and roughly 8 to 12 percent on real-world English audio, and Deepgram Nova-3 reports around 5.26 to 6.84 percent in production conditions. The gap of three to four times between benchmark and production is the figure that should drive a decision, and it is the one usually missing from vendor material.
Why does my voice agent get product names wrong?
Because proper nouns are the weakest part of a general speech model and the most important words in a business conversation. The fix is custom vocabulary or keyword boosting: give the recognizer your product, brand and place names in advance and it stops guessing at them. A misheard name is not only a bad transcript, it also retrieves the wrong passage and produces a confidently wrong spoken answer.
Should I pick a provider on published word error rate?
Use it as a filter, not as the decision. Accuracy belongs to the combination of model, audio conditions, domain vocabulary and speaker, so a provider that leads on clean benchmark English can trail on noisy phone audio in an unfamiliar accent. A 2026 Coval and Gradium benchmark also found no provider leading on both speed and accuracy, so a comparison that ignores latency is incomplete. Test on your own recordings.
Last reviewed 30 July 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.