I've heard people say Parakeet models are better than Whisper models. They weren't lying. In my testing, they're a lot better, and starting with auto-editor 31.3.0, they're easy to use:

```sh
auto-editor whisper video.mp4 ggml-parakeet-tdt-0.6b-v3-q8_0.bin -o video.srt
```

Download a model from [ggml-org/parakeet-GGUF](https://huggingface.co/ggml-org/parakeet-GGUF) on Hugging Face. Auto-Editor selects the backend from the filename, so keep "parakeet" in it. While NVIDIA created Parakeet, you don't need NVIDIA hardware to run it.

## Why Bother?

Parakeet uses a different architecture than Whisper. Whisper is an encoder-decoder transformer that generates text one token at a time. At each decoding step, Parakeet's token-and-duration transducer (TDT) predicts both a token and its duration. That duration lets the decoder skip multiple encoder frames instead of advancing one frame at a time.

Parakeet-TDT has 0.6B parameters—less than half of Whisper large-v3's 1.55B—and its transducer decoding is much cheaper than autoregressive generation. On my machine (macOS 26, M1 Pro MacBook Pro), it transcribes noticeably faster than even the small Whisper models I tried.

The word timings are the other big win. Add `--split-words` to produce one subtitle cue per word:

```sh
auto-editor whisper video.mp4 ggml-parakeet-tdt-0.6b-v3-q8_0.bin --split-words -o video.srt
```

Those tighter cues also produce tighter cuts when editing by spoken content with `--edit word:...`.

## Disadvantages

* There is no `--prompt` equivalent, so you can't nudge it toward proper nouns or jargon.
* Language support is narrower. The v3 model used here supports 25 European languages. NVIDIA's wider Parakeet family includes models for specific languages such as Japanese and Vietnamese, but those models are not available from the linked GGUF repository.

## What About Apple?

Parakeet isn't the only new transcription backend in this release. If you're on macOS 26 or later,  use the magic model name `apple` to select Apple's built-in SpeechAnalyzer:

```sh
auto-editor whisper video.mp4 apple --language en_US -o video.srt
```

SpeechAnalyzer cannot auto-detect the language, so set `--language` to a supported language or locale. Otherwise, Auto-Editor warns and falls back to `en_US`. 

## Benchmarks

I benchmarked all three backends on a [49-minute MIT lecture](https://ocw.mit.edu/courses/6-034-artificial-intelligence-fall-2010/resources/lecture-16-learning-support-vector-machines/) using my M1 Pro MacBook Pro on macOS 26.6. I ran each backend once with the SpeechAnalyzer English model already installed:

| Backend | Time | Speed vs. real time |
| --- | ---: | ---: |
| Parakeet TDT 0.6B v3 Q8 | 1:07.14 | 44.3x |
| Apple SpeechAnalyzer | 5:02.41 | 9.8x |
| Whisper medium.en | 17:19.29 | 2.9x |

Parakeet finished 4.5 times faster than SpeechAnalyzer and 15.5 times faster than Whisper medium.en. SpeechAnalyzer was 3.4 times faster than Whisper medium.en.

The auto-editor CLI will support Parakeet and `apple` models in release 31.3.0.

## Recommendation

For its supported languages, Parakeet will be my default recommendation. In my English-language testing, it is faster than the Whisper models I tried and has better word timings. Reach for `ggml-large-v3` when you need the long tail of languages, translation, or prompting.

```
parakeet >>> apple > whisper
```