For clear speech in a major language, a large Whisper model running privately on your phone or computer is close to online transcription services. Cloud services keep a real edge in a few places. They label speakers reliably, handle very noisy recordings with many people, and offer human-reviewed tiers when every word must be right. In practice, the two things that decide accuracy are the quality of the recording and the size of the model, not whether the model runs in a data center or in your pocket.
What does the research say about Whisper’s accuracy? #
Most private transcription apps today run OpenAI’s Whisper or a close relative, so its published results are the best guide.
- Training scale. The original Whisper was trained on 680,000 hours of multilingual audio from the web. Its authors report that it approaches human accuracy and robustness on many benchmarks without being fine-tuned for them (Radford et al., 2022). The newer large-v3 generation was trained on more than 5 million hours.
- Noise. In the paper’s noise test, models trained only on clean audiobook speech beat Whisper when noise was light. They degraded faster as noise rose, and with pub-style background noise below 10 dB signal-to-noise ratio, Whisper outperformed them all.
- Languages. The same paper notes that accuracy is still poor for many languages and closely tracks how much training data each language had. English and other widely spoken languages do far better than low-resource ones.
- Speed vs size. OpenAI’s large-v3-turbo cuts the decoder from 32 layers to 4, making it “way faster, at the expense of a minor quality degradation” compared with large-v3 (model card).
Many cloud services use Whisper or similar architectures themselves. So on-device versus cloud is often the same kind of model running on different hardware.
Where does accuracy actually differ? #
| Situation | Private, on-device | Online service |
|---|---|---|
| One clear speaker, major language | Very close with a mid-size or large model | Very good |
| Accents | Good with medium, turbo or large models; weak with tiny ones | Good |
| Moderate background noise | Good with larger models and a close mic | Good, sometimes with extra enhancement |
| Several people talking over each other | Words are fine, but no reliable speaker labels | Better, often with named speakers |
| Rare languages | Weaker, depending on training data | Also weaker, varies by provider |
| Names and jargon | Good with a custom vocabulary or prompt | Good with custom vocabulary features |
| Human-verified transcripts | Not available | Available on premium tiers |
Model size is the real variable #
On-device apps have to fit the model into your phone’s memory, and some quietly use the smallest one to feel fast. OpenAI’s Whisper repository lists the sizes:
| Model | Parameters | Relative speed |
|---|---|---|
| tiny | 39 M | ~10x |
| base | 74 M | ~7x |
| small | 244 M | ~4x |
| medium | 769 M | ~2x |
| large | 1,550 M | 1x |
| turbo | 809 M | ~8x |
On clean audio, the gap between small and large is modest. On accented, noisy or fast speech it widens. English-only versions (.en) are more accurate for their size, especially at tiny and base.
An app that lets you pick the model gives you control over this trade-off. Private Transcribe offers five tiers on iPhone and Android. Quick (78 MB) is for instant notes, Balanced (148 MB) is the everyday default, and Accurate (190 MB) handles noisy rooms and accents. Professional (539 MB) is for meetings and interviews, and Best (574 MB) is Whisper large-v3-turbo. All of them run on the phone, so choosing more accuracy costs time, not privacy.
The failure modes are the same in the cloud #
Whisper’s authors describe errors that aren’t about hearing: repeat loops, dropping the first or last words of a segment, and occasionally complete hallucination unrelated to the audio. An independent 2024 study, Careless Whisper, found roughly 1% of Whisper transcriptions contained whole phrases that weren’t in the audio, more often in recordings with long pauses. Those errors come from the model, so they apply wherever it runs, including cloud services built on it.
The practical defence is the same everywhere. Trim long silences, use a larger model, and check anything you’ll quote against the recording. Timestamps make that quick. Our explainer on why offline speech-to-text accuracy drops with noise goes deeper on the noise side.
How to measure accuracy on your own audio #
Published benchmarks use clean test sets. Your recordings aren’t clean test sets, so a 20-minute check is worth more than any leaderboard.
- Pick a representative 3 to 5 minute clip, with your usual mic, room and speakers.
- Type an exact reference transcript by listening carefully.
- Run the clip through each tool you’re considering, using its best setting.
- Count errors: words substituted, missed or added. Word error rate (WER) is that total divided by the number of words in your reference.
- Note what kind of errors each tool makes. A tool that misses names is easier to fix with a vocabulary list than one that drops whole sentences.
For most day-to-day notes, anything under about 10% WER reads fine after a quick skim. For quotes, legal or medical use, you’ll proofread whatever you use.
When is an online service the better choice? #
- Court filings, publication and compliance records, where a human-verified transcript is worth paying for.
- Large meetings where you need every speaker named automatically.
- Huge archives where your own hardware would take weeks.
- Team workflows built around shared, searchable transcripts.
If the audio is confidential, weigh those benefits against uploading it. The privacy side is covered in on-device vs cloud transcription. For journalists specifically, see private transcription apps for journalists.
How to get cloud-level accuracy offline #
- Record better. Get the mic close to the speaker. This beats any setting. See how to minimize background noise.
- Use the largest model your device handles well for anything important.
- Pin the spoken language instead of relying on auto-detect.
- Add custom vocabulary for names and jargon. Private Transcribe has a vocabulary list in Settings, and whisper.cpp takes
--prompt. - Trim long silences and music to reduce invented text.
- Proofread against the audio using timestamps.
Frequently asked questions #
Is on-device transcription as accurate as Otter or Rev? #
For clear, single-speaker audio in a major language, a large on-device Whisper model is in the same range as automated cloud services. Cloud services do better at naming speakers in group recordings, and human transcription tiers are more accurate than any automated option. Test on your own audio before deciding.
Does a smaller file size mean worse accuracy? #
Smaller models (fewer parameters) are less accurate, especially on difficult audio. Quantization is different. It stores the same model more compactly and usually has only a small effect on accuracy, which is why phone apps can offer large models at a few hundred megabytes.
Why does my transcript contain words nobody said? #
Whisper-based models sometimes produce text during long silences, music or heavy noise, because they predict words from context. It happens in cloud services built on the same models too. Trimming silence and using a larger model reduce it.
Which is more accurate for non-English languages? #
It depends on the language. Whisper’s multilingual models are strong in widely spoken languages and weaker in low-resource ones. Some cloud providers train extra data for specific languages, so for a less common language, compare a few tools on a sample.
Is private transcription accurate enough for professional work? #
For notes, drafts, research and internal meetings, yes, with a large model and good audio. For verbatim legal records or published quotes, use it for the first draft and proofread against the recording, just as you would with any automated transcript.