OpenAI Whisper Alternatives: 12 Speech-to-Text Options
November 17, 2026

OpenAI Whisper is two things, and the right alternative depends on which one you use. If you call whisper-1 through the OpenAI API, you have to move: OpenAI shuts it down on February 26, 2027, and points you to gpt-transcribe or gpt-live-transcribe. If you run the open-weight Whisper model yourself, nothing is shutting down, but newer open models beat it on speed or size: NVIDIA Parakeet and Canary for European languages, Mistral's Voxtral Realtime for streaming, and Moonshine for phones and edge devices. If you just want Whisper faster, swap the runtime to faster-whisper or whisper.cpp. If you'd rather not host anything, Deepgram, AssemblyAI, ElevenLabs Scribe and Google Chirp 3 are the hosted options.
This guide is for developers who built on Whisper, the model or the API, and want to know what to switch to in late 2026. It covers models and APIs, not desktop apps.
Quick Comparison
| Option | Type | Languages | Streaming | License or billing |
|---|---|---|---|---|
| OpenAI gpt-transcribe / gpt-live-transcribe | Hosted API | Multilingual | Yes | Usage-based |
| NVIDIA Parakeet TDT 0.6B v3 | Open weights | 25 European | Not covered here | CC-BY-4.0 |
| NVIDIA Canary 1B v2 | Open weights | 25 European, plus translation | Not covered here | CC-BY-4.0 |
| Mistral Voxtral Mini 4B Realtime | Open weights (API on Voxtral 3) | 13 | Yes | Apache 2.0 |
| Moonshine | Open weights, on-device | 10 for speech-to-text | Yes | MIT (one legacy exception) |
| faster-whisper | Whisper runtime | Same as Whisper | Not covered here | MIT |
| whisper.cpp | Whisper runtime | Same as Whisper | Not covered here | MIT |
| Distil-Whisper | Smaller Whisper | English only | Not covered here | MIT |
| Deepgram Nova-3 / Flux | Hosted API | Broad multilingual | Yes | Usage-based |
| AssemblyAI STT-3.7 | Hosted API | 32 (99 on Universal-2) | Yes | Usage-based |
| ElevenLabs Scribe v3 | Hosted API | 90+ | Yes (Scribe v3 Live) | Usage-based |
| Google Chirp 3 | Hosted API | Many locales, about 30 GA | Yes | Usage-based |
What's Happening to Whisper in 2026
Start with the API, because that's the part with a deadline. OpenAI's deprecations page lists whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize with a notice date of August 26, 2026 and a shutdown date of February 26, 2027. The /v1/audio/translations endpoint goes on the same day. The replacements are gpt-transcribe or gpt-live-transcribe, and /v1/realtime/translations for translation.
So "switch to gpt-4o-transcribe" is no longer the answer. It's on the same clock as Whisper.
The open model is a different story. The openai/whisper repo still ships the code and weights under the MIT license, in five sizes from tiny (39M parameters) to large (1550M), plus turbo. Turbo, an 809M-parameter optimized version of large-v3, is faster with what OpenAI calls minimal accuracy loss, but it isn't trained for translation. If you self-host, you can keep running it as long as you like.
Why do teams look elsewhere anyway?
- Speed and cost on GPUs. Whisper large is slow next to newer models of similar accuracy.
- No live streaming. Whisper works on 30-second windows. Real-time use takes workarounds.
- Hallucinations on silence. Long pauses and music can sometimes produce text nobody said.
- No built-in diarization. Speaker labels need a second model.
- Size on very small devices. Even the tiny model is large for a microcontroller.
OpenAI's Own Replacement: gpt-transcribe
Best for: teams already on the OpenAI API who want the shortest migration.
OpenAI's speech-to-text guide now recommends gpt-transcribe for recorded audio and gpt-live-transcribe through the Realtime API for live audio. gpt-transcribe streams text deltas when you set stream=true, takes a languages list in place of the old single language field, and accepts files up to 25 MB in the same formats as before.
Watch for three gaps when you migrate. verbose_json timestamps are documented only for whisper-1. Speaker labels come from gpt-4o-transcribe-diarize, which is itself deprecated. And gpt-transcribe keeps the original language, so English translation moves to the separate realtime translations endpoint.
Where it beats Whisper: no hosting, streaming output, and one API key with the rest of your OpenAI stack.
Where it trails Whisper: no open weights, and the timestamp and diarization story is thinner than the hosted specialists below.
Open-Weight Whisper Alternatives You Can Self-Host
NVIDIA Parakeet TDT 0.6B v3: fast batch transcription for European languages
Parakeet TDT 0.6B v3 is a 600M-parameter model that NVIDIA expanded from English to 25 European languages. It returns word-level and segment-level timestamps, and NVIDIA says it handles audio up to 24 minutes with full attention on an A100 80GB, or up to 3 hours with local attention. The license is CC-BY-4.0, which allows commercial use with attribution.
Where it beats Whisper: a smaller model with word-level timestamps, built for high-throughput batch work through NVIDIA NeMo.
Where it trails Whisper: 25 European languages, not Whisper's long multilingual list. If you transcribe Japanese, Hindi or Arabic, look elsewhere.
NVIDIA Canary 1B v2: transcription plus translation
Canary 1B v2 is the bigger sibling at 978M parameters, also CC-BY-4.0. It transcribes the same 25 European languages and translates in both directions: English to 24 languages, and 24 languages to English. Transcripts get word-level and segment-level timestamps, and translations get segment-level ones.
Where it beats Whisper: translation out of English, not just into it, which Whisper never did.
Where it trails Whisper: the same European-only coverage as Parakeet, and a heavier model.
Mistral Voxtral Realtime: open weights built for streaming
The open-weight model is Voxtral Mini 4B Realtime, released under Apache 2.0. Its model card describes a roughly 4B-parameter streaming model for 13 languages with a configurable delay, recommends a 480ms setting, and claims accuracy close to offline open-source models at that delay. Mistral recommends serving it with vLLM.
Mistral's hosted API has since moved to a third generation. Its speech-to-text docs list Voxtral Mini Transcribe 3 and Voxtral Small Transcribe 3 for recorded audio, with speaker labels and segment timestamps, and Voxtral Mini Transcribe Realtime 3 for live audio with end-of-turn detection. The docs don't mention open weights for those models.
Where it beats Whisper: real streaming in an Apache 2.0 model, plus a hosted API with speaker labels if you don't want to run it.
Where it trails Whisper: 13 languages, and the open model is a generation behind Mistral's hosted Voxtral 3 models.
Moonshine: speech-to-text on phones and edge devices
Moonshine, now under the moonshine-ai org, is an on-device speech toolkit with models trained from scratch. Its README describes a range from models it says beat Whisper Large V3 on accuracy down to tiny 1MB models, all optimized for live streaming. Speech-to-text covers English, Spanish, German, Mandarin, Japanese, Korean, Vietnamese, Tagalog, Ukrainian and Arabic, and it runs in Python, JavaScript and WebAssembly, iOS, Android, macOS, Linux, Windows and Raspberry Pi.
Check the license per model. The code and most models are MIT, but the legacy non-streaming models for non-English languages use a non-commercial community license.
Where it beats Whisper: size, streaming and on-device support with no account or API key. If you're putting speech on a phone, pair it with our guide to running an LLM on Android.
Where it trails Whisper: fewer languages, and the accuracy claims are the vendor's own.
Faster Ways to Run Whisper Itself
Sometimes you don't need a new model. You need Whisper to stop being slow. These keep Whisper's weights and change the engine.
| Project | What it changes | Best for | License |
|---|---|---|---|
| faster-whisper | CTranslate2 runtime, 8-bit quantization | Python servers and GPU batch jobs | MIT |
| whisper.cpp | C/C++ port, Metal, CUDA, Vulkan, Core ML | Macs, desktops, phones, browsers | MIT |
| Distil-Whisper | Distilled, smaller English checkpoints | English-only workloads | MIT |
faster-whisper reimplements Whisper on CTranslate2. Its README says it's up to 4 times faster than openai/whisper for the same accuracy while using less memory, and 8-bit quantization cuts more on CPU and GPU. It loads large-v3, turbo and Distil-Whisper checkpoints.
whisper.cpp is a plain C/C++ port with no dependencies. On Apple Silicon it runs fully on the GPU through Metal, and the encoder can run on the Neural Engine through Core ML. It also supports CUDA, Vulkan, iOS, Android and WebAssembly, plus quantized models that need less memory. If you want a desktop app on top of Whisper instead of a library, our MacWhisper alternatives guide covers those.
Distil-Whisper distills Whisper into smaller checkpoints. The project says distil-large-v3 is 6 times faster and 49% smaller than large-v3, within 1% word error rate on out-of-distribution sets. A distil-large-v3.5 checkpoint is also published on Hugging Face, though the project still recommends distil-large-v3. All of it is English only. For multilingual audio, the project points you back to Whisper turbo.
Hosted Speech-to-Text APIs
If you used the Whisper API because you didn't want to host anything, these are the closest swaps. Each has a richer feature set than whisper-1 ever had.
Deepgram Nova-3 and Flux
Deepgram's models overview lists Flux for voice agents, with built-in end-of-turn detection in English and a ten-language multilingual version. Nova-3 is the general model for meetings, captions and noisy audio, in batch or streaming, with medical and pharma variants. Nova-2 covers languages Nova-3 doesn't yet. Deepgram also hosts Whisper itself as Whisper Cloud, though its docs say those models scale worse than its own. For a deeper look, see our Deepgram alternatives guide.
AssemblyAI STT-3.7
AssemblyAI's lineup has changed again. Its models page now lists AssemblyAI-STT-3.7 as the flagship for recorded audio, with 32 languages, code switching, and keyterm prompting of up to 1,000 words. Universal-3.5 Pro is marked deprecated. Streaming has two 3.7 variants, one for voice agents and one for notetakers with real-time speaker labels, and Universal-2 stays on for 99 languages.
ElevenLabs Scribe v3
The ElevenLabs speech-to-text docs list Scribe v3 for recorded audio and Scribe v3 Live for real time, across 90+ languages. Scribe v3 handles diarization for up to 32 speakers, word-level timestamps, audio events like laughter, and keyterm prompting. ElevenLabs reports 27% fewer word errors than Scribe v2 on FLEURS, a vendor figure.
Google Chirp 3
Chirp 3 is Google's current speech model, available only on the Speech-to-Text V2 API. It supports streaming, synchronous requests for clips under a minute, and batch jobs for longer files. About 30 locales are GA and the rest are in preview. Diarization and automatic language detection are GA, though diarization covers a smaller set of languages.
How to Choose a Whisper Alternative
| Your situation | Look at | Why |
|---|---|---|
You call whisper-1 and want the least work | gpt-transcribe | Same endpoint, OpenAI's named replacement |
| You need speaker labels and timestamps | AssemblyAI, ElevenLabs Scribe, Deepgram | Built into the API |
| You self-host and want more speed | faster-whisper, whisper.cpp | Same model, faster engine |
| European languages at high volume | Parakeet, Canary | Small, fast, timestamps included |
| Live streaming with open weights | Voxtral Realtime, Moonshine | Built for streaming |
| On a phone or edge device | Moonshine, whisper.cpp | Small models, mobile builds |
| English only, on a budget GPU | Distil-Whisper | Smaller and faster checkpoints |
| You're on Google Cloud | Chirp 3 | One bill, existing compliance |
Before you switch, run a test on your own audio. Take 30 clips that look like production, including accents, crosstalk, background noise and the product names your users say. Score word error rate on the words that matter, not just overall. Then check what happens on long silences, since that's where Whisper tends to invent text.
When Whisper Is Still the Right Call
Keep the open model if:
- You need many languages from one open model. Whisper's multilingual coverage is still broader than most open alternatives.
- Your pipeline already works. Self-hosted Whisper isn't going away, and faster-whisper or whisper.cpp can fix speed without a model change.
- You need permissive licensing everywhere. MIT on code and weights is simple to approve.
- You want the biggest community. Most open speech tools support Whisper first.
Where Taku Fits
Every option here is a building block for developers wiring speech into a product. Many people never build that pipeline; they just want AI apps someone else already put together. Taku isn't a speech-to-text API, and it doesn't replace Whisper or any model here. That's where Taku comes in.
Taku Apps and Sites, built by the best AI minds. Find one, run it, and keep it in your Mac Dock. Taku is in Beta, and the Mac app is available now.
Key Points
- OpenAI shuts down
whisper-1,gpt-4o-transcribe,gpt-4o-mini-transcribeand the translations endpoint on February 26, 2027. The replacements aregpt-transcribeandgpt-live-transcribe. - The open-weight Whisper model stays MIT-licensed and keeps working for self-hosters.
- NVIDIA Parakeet and Canary are fast CC-BY-4.0 models for 25 European languages. Canary adds translation.
- Voxtral Realtime is Apache 2.0 and built for streaming. Moonshine targets phones and edge devices.
- faster-whisper, whisper.cpp and Distil-Whisper speed up Whisper without leaving it.
- Hosted alternatives are Deepgram Nova-3 and Flux, AssemblyAI STT-3.7, ElevenLabs Scribe v3 and Google Chirp 3.
FAQ
Is OpenAI Whisper being discontinued?
The API model is. OpenAI's deprecations page lists whisper-1 for shutdown on February 26, 2027. The open-source Whisper model and its MIT-licensed weights on GitHub aren't affected.
What replaces whisper-1 in the OpenAI API?
OpenAI names gpt-transcribe for recorded audio and gpt-live-transcribe for live audio. gpt-4o-transcribe and gpt-4o-mini-transcribe are deprecated on the same date as whisper-1, so don't migrate to them.
What is the best open-source alternative to Whisper?
For European languages, NVIDIA Parakeet TDT 0.6B v3 is fast and accurate, and Canary 1B v2 adds translation. For streaming, try Voxtral Realtime. For phones and edge devices, try Moonshine. Check each license: Parakeet and Canary are CC-BY-4.0, Voxtral Realtime is Apache 2.0, and Moonshine is mostly MIT.
Is faster-whisper better than Whisper?
It's the same model on a faster engine. The project says it runs up to 4 times faster than openai/whisper at the same accuracy with less memory. Accuracy doesn't change, because the weights are Whisper's.
Can I run a Whisper alternative locally on a Mac?
Yes. whisper.cpp runs Whisper on Apple Silicon through Metal and Core ML, and Moonshine supports macOS. Both work offline once the model is downloaded.
Which Whisper alternative supports speaker diarization?
AssemblyAI, ElevenLabs Scribe v3, Deepgram, Google Chirp 3 and Mistral's Voxtral Transcribe 3 models all label speakers. Self-hosted Whisper doesn't, so you'd need a separate diarization model.