There is no single best speech to text model. The best voice to text model for you depends on four things: whether audio may leave your computer, which languages you speak, whether you need live results or file transcripts, and whether you need extras such as speaker labels. If audio must stay on your Mac, look at open models such as Whisper and NVIDIA's Parakeet. If a cloud service is acceptable, OpenAI, Google, Deepgram, AssemblyAI and ElevenLabs all publish current models with features the open ones lack. Then test the shortlist on your own voice.
Publisher note: SuprFlow publishes this guide and is one of the products discussed. Competitor details come from each company's own pages, checked on 1 October 2026. Prices and features change; confirm before you buy.
We have not run a controlled benchmark of any model here. Where we quote error rates, they describe the test datasets their publishers used, not your voice, accent, microphone or vocabulary.
Which model fits which need?
| If you need | Look at first | Why, per the publisher |
|---|---|---|
| Audio that never goes to a speech server | Whisper (any size), Parakeet | Open weights you can run locally (OpenAI README, NVIDIA cards) |
| Many languages, run locally | Whisper large-v3 or turbo; Qwen3-ASR | Whisper's card lists 99 language codes; Qwen3-ASR covers "52 languages and dialects" |
| English only, run locally | Parakeet TDT 0.6B v2, Whisper .en models | English-specific models from NVIDIA and OpenAI |
| European languages, run locally | Parakeet TDT 0.6B v3 | 25 European languages per its card |
| Speaker labels on recordings | AssemblyAI, ElevenLabs Scribe v2, OpenAI gpt-4o-transcribe-diarize, Google Chirp 3 | Each documents diarization |
| Live streaming from an app or call | Deepgram Flux or Nova-3, ElevenLabs Scribe v2 Realtime, AssemblyAI streaming models, OpenAI Realtime | Each documents streaming |
| Live dictation into any Mac app | A dictation app built on a local model | The model alone does not type into your apps |
Open models you can run yourself
| Model | Publisher | Size | Languages | Licence |
|---|---|---|---|---|
| Whisper large-v3 | OpenAI | 1550M parameters | 99 language codes | MIT per OpenAI's README (Hugging Face tag reads apache-2.0) |
| Whisper large-v3-turbo | OpenAI | 809M | Same list; not trained for translation | MIT |
| Whisper tiny to medium | OpenAI | 39M to 769M | Multilingual, plus English-only .en versions | MIT |
| Parakeet TDT 0.6B v2 | NVIDIA | 600M | English | CC-BY-4.0 |
| Parakeet TDT 0.6B v3 | NVIDIA | 600M | 25 European languages | CC-BY-4.0 |
| Canary-Qwen-2.5B | NVIDIA | 2.5B | English | CC-BY-4.0 |
| Qwen3-ASR-1.7B | Qwen | 1.7B | 30 languages and 22 Chinese dialects | Apache 2.0 |
Sources: the Whisper README and large-v3 card, the Parakeet v2 and v3 cards, the Canary-Qwen-2.5B card and the Qwen3-ASR-1.7B card. Size and licence alone do not tell you whether a model runs well on a Mac. Whisper has Mac-focused runtimes (whisper.cpp and WhisperKit); check before you pick another model that a tool exists to run it on your hardware. Whisper vs faster-whisper vs WhisperX covers Whisper's options and Parakeet vs Whisper compares those two families in depth.
Cloud speech to text AI from the major providers
Each row uses only the provider's own model page. We have not compared prices here; check each pricing page.
| Provider | Current models it documents | What it says they are for |
|---|---|---|
| OpenAI | gpt-transcribe, gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-4o-transcribe-diarize, whisper-1, plus a realtime option | "Start with gpt-transcribe. This is the recommended model for transcribing recorded speech in its original language." Files can be up to 25 MB |
| Google Cloud | Chirp 3 | "the latest generation of Google's multilingual Automatic Speech Recognition (ASR)-specific generative models"; 40 languages at general availability, more in preview; diarization and speech adaptation |
| Deepgram | Flux, Nova-3, Nova-2 | Flux for "real-time agents, customer support bots, and interactive, turn-based experiences"; Nova-3 for "meetings, event captioning, multi-speaker, multilingual, noisy, or far-field audio" |
| AssemblyAI | Universal-3.5 Pro, Universal-2, streaming models | Universal-3.5 Pro: "Highest accuracy, fastest model", 18 languages; Universal-2 for 99 languages |
| ElevenLabs | Scribe v2, Scribe v2 Realtime | Scribe v2 lists word-level timestamps, speaker diarization and "Keyterm prompting"; Realtime says it returns partial transcripts in "~150 milliseconds" |
A cloud model means your audio is sent to that provider. That is fine for many jobs and unacceptable for some. Read the provider's data terms, not only its model page.
What does the Open ASR Leaderboard tell you?
The Hugging Face Open ASR Leaderboard is the most widely used public comparison. Its About tab says models are "ranked based on their Average WER," and that speed is reported as RTFx. It includes both open models and paid APIs. In the results file for its 25 September 2026 update, proprietary entries include AssemblyAI Universal-3.5 Pro, ElevenLabs Scribe v2, Microsoft Azure Speech and Speechmatics, alongside open models from NVIDIA, OpenAI, Qwen, IBM and others.
Three limits before you treat it as a ranking:
- Not everyone is listed. We found no entry for OpenAI's
gpt-transcribe, Google's Chirp 3 or Deepgram in that results file. Absence is not a verdict. - It changes often. Its version notes show updates on 11, 19 and 25 September 2026. Read the live page rather than any copy, including ours.
- It is mostly English. The main table uses English datasets such as audiobooks, meetings, earnings calls and parliament speeches. Separate tabs cover a few other languages. Text is normalised before scoring, so punctuation and capitals are not tested.
We do not reproduce its order here. Speech to text accuracy and word error rate explains how to read it.
Where does SuprFlow fit?
SuprFlow is a Mac dictation app, not a model. Version 1.0.0 runs a WhisperKit (Whisper) model on the Mac, with Small and Medium options in multilingual and English-only variants. It does not offer Parakeet, turbo, large-v3 or any cloud speech model, and it does not transcribe audio files or label speakers. If you need those, one of the options above is the better choice.
How do you pick between your top two?
- Write a fictional script with a name, a date, an amount and a negation: "Mateus Oyelaran moved the 2 April call to 10:45. Revenue was 1.8 million, not 18. Do not copy Ines."
- Record it once in a quiet room and once with background noise, using the microphone you normally use.
- Run both models on the same recordings.
- Count wrong names, numbers and negations, not just total words.
- If you need another language, repeat in that language.
Pick the model that makes fewer errors that matter to you.
Questions people ask
What is the most accurate speech to text model?
It depends on the audio and language, and public rankings change week to week. Read the Open ASR Leaderboard for current English results, then test your top two on your own recordings. We have not run a benchmark of our own.
What is the best free speech to text model?
The open models above are free to download: Whisper under MIT, Parakeet and Canary-Qwen under CC-BY-4.0, Qwen3-ASR under Apache 2.0. Running them costs your own computer time. Check each licence's attribution terms.
Is Whisper still a good choice in 2026?
Yes, for local and multilingual use. It has the widest language list of the open models here and the most tools for running on a Mac. In the leaderboard's results file, newer models such as Parakeet TDT 0.6B v2 show lower error rates than Whisper large-v3 on every main English dataset, so compare if you only need English.
Which speech to text model works offline on a Mac?
Any open model you can run locally, such as Whisper through whisper.cpp or WhisperKit. Cloud models need a connection. Voice to text without internet on Mac lists the practical options.
Related guides
- Parakeet vs Whisper
- Whisper model sizes compared
- Speech to text accuracy and word error rate
- On-device speech recognition explained
- Voice to text without internet on Mac
- Best dictation app for Mac
Sources
- OpenAI — whisper GitHub repository and README
- OpenAI — whisper-large-v3 model card
- whisper-large-v3-turbo model card
- NVIDIA — parakeet-tdt-0.6b-v2
- parakeet-tdt-0.6b-v3
- canary-qwen-2.5b
- Qwen — Qwen3-ASR-1.7B model card
- OpenAI — Speech to text guide
- Google Cloud — Chirp 3 model documentation
- Deepgram — Models and languages overview
- AssemblyAI — Models
- ElevenLabs — Models overview
- Hugging Face — Open ASR Leaderboard
- English results dataset