# Automatic speech recognition Canonical page: https://flit.fyi/glossary/automatic-speech-recognition [Glossary](https://flit.fyi/glossary) / [How it hears you](https://flit.fyi/glossary#recognition) Also: ASR Updated October 16, 2026 Definition Automatic speech recognition (ASR) is the machine-learning field, and the class of models, that turns audio of human speech into text. ASR models are the engine inside every dictation and transcription product. Modern ASR models are neural networks trained on large amounts of recorded speech. Well-known families include Whisper (OpenAI, open source), Parakeet (NVIDIA, open source), and the proprietary engines inside cloud services from Google, Apple, and others. ASR quality is usually reported as word error rate (WER) on benchmark audio. Real-world quality also depends on your microphone, room, accent, and vocabulary, which is why per-user dictionaries and domain vocabularies matter as much as the headline model. ## How the models are built An ASR model is trained on pairs of audio and the correct transcript, hundreds of thousands of hours of them, until it can predict the text for audio it has never heard. Older systems separated the acoustic model (sounds to phonemes) from the language model (phonemes to likely words). Current systems are end-to-end: one network maps audio directly to text tokens, which is simpler, more accurate, and easier to run on consumer hardware. Two architectures dominate today. Encoder-decoder models, like Whisper, read a whole chunk of audio and then write the transcript, which makes them accurate and batch-oriented. Transducer models, like Parakeet, emit tokens as audio arrives, which makes them streamable. ## What decides accuracy in practice Model size and training data set the ceiling. Everything below it is yours: microphone quality, distance from it, room echo, background noise, accent, and vocabulary. Names, jargon, and product terms are where every model struggles, because they are rare in training data, which is why a personal dictionary and a context-aware correction step tend to improve day-to-day accuracy more than switching to a bigger model. ## Open models changed the category Until 2022, competitive ASR lived behind cloud APIs. Whisper’s open release, followed by Parakeet and others, put state-of-the-art recognition in anyone’s hands, and Apple silicon gave laptops the compute to run it. That is the reason private, local dictation went from a compromise to a default in a few years. ## How to compare two ASR models on your own voice Benchmark scores describe benchmark audio. The only way to know how a model handles your voice, your room, and your vocabulary is to run it on a recording of you. The open Whisper command-line tool makes this a short job in Terminal. It needs Python and ffmpeg installed first. 1. Record one or two minutes of yourself speaking the way you normally dictate, including a few names and technical terms, and save it as an audio file. 2. Install Whisper in Terminal with: pip install -U openai-whisper 3. Transcribe the file with a small model: whisper myclip.m4a --model small --output_format txt 4. Rename the text file it wrote, then run the same command with --model turbo. 5. Read the two transcripts against what you actually said and count the wrong, missing, and extra words in each. 6. Divide the error count by the number of words you spoke. That is your personal word error rate for each model. ## What an ASR model returns besides words Older recognizers returned a flat run of lowercase words. Current open models return more, and knowing what is available tells you what a product built on them can do without extra work. - Punctuation and capitalization. Parakeet TDT 0.6B lists both as built-in features. - Timestamps. Parakeet returns word-level and segment-level timestamps, and Whisper has a word_timestamps option. - Subtitle files. The Whisper command-line tool can write txt, vtt, srt, tsv, and json output. - The language. Whisper and Parakeet TDT 0.6B v3 both detect the spoken language when you do not specify it. - A translation. Whisper’s multilingual models can translate speech in other languages into English text. ## Is ASR accurate? Why one model has many scores There is no single accuracy number for an ASR model. NVIDIA reports an average word error rate of 6.34% for Parakeet TDT 0.6B v3 across the eight English test sets of the Hugging Face Open ASR Leaderboard, but the individual sets range from 1.93% on LibriSpeech test-clean to 11.42% on Earnings-22, which is recordings of company earnings calls. Language matters as much as audio type. The Whisper model card says the models show strong results in about 10 languages, and that accuracy is lower on languages with less training data. When you read a score, check which test set and which language it was measured on before you compare it with another. In Flit Flit runs more than one speech engine, all of them on your Mac, and switches between them by language. [How Flit processes speech](https://flit.fyi/privacy) - [Which languages stream and which do not](https://flit.fyi/#languages) - [The dictionary that fixes what the engine mangles](https://flit.fyi/features#dictionary) - [Where the models run](https://flit.fyi/privacy) ## Questions people ask. ### What is the best ASR model for dictation? For live dictation, a streaming-capable model such as Parakeet, because words appear as you speak. For the widest language coverage, a model trained on many languages, such as Whisper with 99 or Meta’s Omnilingual ASR with more than 1,600. Most good tools use more than one and pick by language and job. On Apple silicon all of them run locally. ### Is ASR the same as speech-to-text? ASR is the technical name for the models and the research field. Speech-to-text is the plain-language name for what they do. On a product page the two are interchangeable. ### Why does ASR get my names wrong? Because a model has only seen a name as often as it appeared in training audio, which for most names and product terms is rarely. It writes the most probable spelling of the sounds. A dictionary of your own terms, and correction against names it can see on screen or in your project, fix this better than any model swap. ### What does ASR stand for? ASR stands for automatic speech recognition. It is the technology that turns recorded or live speech into text. ### Can I run an ASR model on my own computer? Yes. Open models such as Whisper and Parakeet can be downloaded and run locally. Whisper installs with pip and runs from Terminal on a Mac. ### Do ASR models add punctuation? Many current models do. NVIDIA lists automatic punctuation and capitalization as features of Parakeet TDT 0.6B, and Whisper also writes punctuated text. ## Compared on it - [Flit vs FluidVoice](https://flit.fyi/alternatives/fluidvoice) ## Related terms - [Whisper](https://flit.fyi/glossary/whisper) - [Parakeet](https://flit.fyi/glossary/parakeet) - [Word error rate](https://flit.fyi/glossary/word-error-rate) - [Hallucination](https://flit.fyi/glossary/hallucination) - [Speech-to-text](https://flit.fyi/glossary/speech-to-text) ## Your voice was always faster. 7 days free, then $19 USD once. No account, no subscription, nothing uploaded. [Download for Mac](https://flit.fyi/download) [Buy now · $19 USD](https://flit.fyi/pricing#personal) macOS 14 or later · Apple silicon