Speech-to-text model comparison and how to choose
We measured the 10 speech-to-text models available in this tool on the same audio in 5 languages. The error rate is the share of the transcript that was recognized incorrectly, and the processing time is an estimate of how long 1 hour of audio takes. If you are not sure which model to choose, start with "Recommendations by language and device".
Measured on: October 29, 2026
Recommendations by language and device
Japanese
On a computer with WebGPU, Haas-ja v1.0 (710 MB) is the most accurate (5.2% error rate). It also takes about a third of the time SenseVoice Small does. It is the default in Chrome and Edge on such computers (except Macs with Apple silicon). On phones and computers without WebGPU, choose the default, SenseVoice Small (240 MB). It is about as accurate as Whisper Large v3 Turbo (1.5 GB) and takes less than half the time.
Chinese and Korean
For Chinese, the default SenseVoice Small is more accurate than any Whisper model. For Korean, on a computer with WebGPU, Whisper Large v3 Turbo (5.0% error rate) and Large v3 (4.6%) are more accurate than SenseVoice Small (7.9%). For both languages, Qwen3-ASR 1.7B on a computer with WebGPU is as accurate as SenseVoice Small or better.
English
On a computer with WebGPU, Whisper Large v3 Turbo (5.8% error rate) or Medium (5.7%) is accurate enough. The most accurate is Qwen3-ASR 1.7B (4.3%), but it downloads about 5.2 GB the first time and uses about 5 GB of GPU memory. On phones, choose SenseVoice Small (9.0%).
Spanish, German and other languages
On a computer with WebGPU, choose Whisper Large v3 Turbo. For Spanish, Turbo (4.4% error rate) is more accurate than Large v3 (6.2%) and also faster. The most accurate for Spanish is Qwen3-ASR 1.7B (4.1%). On phones and computers without WebGPU, choose Whisper Small or larger.
Phones
For Japanese, Chinese, Korean and English, choose the default, SenseVoice Small. Phones run a lighter (8-bit) version of Whisper, so Tiny's error rate for Japanese exceeds 70%. For other languages, Whisper Small is the most accurate, but processing takes about as long as the audio or longer. For long audio, Whisper Base is also an option.
WebGPU is a way for the browser to use the GPU for computation. It is available in desktop Chrome, Edge and other browsers.
Open the transcription toolError rates by language
The error rate is the share of the transcript that differs from the correct text. Lower is more accurate. Japanese, Chinese and Korean are counted per character, and English and Spanish per word. The model with the fewest errors in each language is highlighted. The table shows results on a computer with WebGPU.
| Model | Japanese | English | Spanish | Korean | Chinese |
|---|---|---|---|---|---|
| Haas-ja v1.0 | 5.2% | — | — | — | — |
| SenseVoice Small | 8.6% | 9.0% | — | 7.9% | 7.9% |
| Whisper Tiny | 44.4% | 16.0% | 21.1% | 20.4% | 30.1% |
| Whisper Base | 26.3% | 11.5% | 13.2% | 13.7% | 24.0% |
| Whisper Small | 14.1% | 7.3% | 7.6% | 7.2% | 14.5% |
| Whisper Medium | 8.7% | 5.7% | 5.3% | 5.7% | 9.8% |
| Whisper Large v3 Turbo | 8.9% | 5.8% | 4.4% | 5.0% | 9.9% |
| Whisper Large v3 | 6.9% | 5.6% | 6.2% | 4.6% | 11.5% |
| Qwen3-ASR 1.7B | 8.0% | 4.3% | 4.1% | 5.1% | 7.7% |
| Voxtral Mini 4B | 6.1% | — | — | — | — |
"—" means the model cannot be selected for that language in this tool. These results are for read-aloud speech, so every model makes more errors on noisy recordings or conversations between several people.
Processing time and download size
The processing time is an estimate of how long it takes to transcribe 1 hour of audio on our test computer (AMD Ryzen 7 8745H, Radeon 780M integrated graphics). The download size is downloaded only the first time; after that, the model saved in the browser is used.
| Model | Time for 1 hour of audio | First-time download | Requirements |
|---|---|---|---|
| Haas-ja v1.0 | about 2 min | 〜710 MB | Computer with WebGPU (uses about 1 GB of GPU memory) |
| SenseVoice Small | about 7 min | 〜240 MB | Computers and phones (runs on the CPU) |
| Whisper Tiny | about 5 min | 〜145 MB | Computers and phones |
| Whisper Base | about 6 min | 〜280 MB | Computers and phones |
| Whisper Small | about 10 min | 〜925 MB | Computers and phones |
| Whisper Medium | about 19 min | 〜650 MB | Computer with WebGPU recommended (on the CPU, 1 hour of audio takes several hours) |
| Whisper Large v3 Turbo | about 16 min | 〜1.5 GB | Computer with WebGPU and a GPU that supports 16-bit floats (fp16) recommended (without it, it runs on the CPU and takes several hours) |
| Whisper Large v3 | about 23 min | 〜2.0 GB | Computer with WebGPU and a GPU that supports 16-bit floats (fp16) |
| Qwen3-ASR 1.7B | about 19 min | 〜5.2 GB | Computer with WebGPU (uses about 5 GB of GPU memory) |
| Voxtral Mini 4B | about 50 min | 〜2.9 GB | Computer with WebGPU and a GPU that supports 16-bit floats (fp16) (uses about 5 GB of GPU memory; integrated graphics need 16 GB of RAM or more) |
It is faster on computers with a dedicated GPU such as GeForce, and slower on older computers. SenseVoice Small does not use the GPU; it runs on the CPU.
On phones and computers without WebGPU
On phones and computers without WebGPU, a lighter (8-bit) version of Whisper runs on the CPU. Error rates are higher than on a computer's GPU, and Tiny in particular gets much worse.
| Model | Japanese | English | Spanish | Korean | Chinese | Time for 1 hour of audio | First-time download |
|---|---|---|---|---|---|---|---|
| SenseVoice Small | 8.6% | 9.0% | — | 7.9% | 7.9% | about 7 min | 〜240 MB |
| Whisper Tiny | 70.5% | 28.3% | 31.5% | 30.2% | 51.0% | about 10 min | 〜40 MB |
| Whisper Base | 31.2% | 12.1% | 14.0% | 15.2% | 25.1% | about 15 min | 〜75 MB |
| Whisper Small | 14.8% | 7.6% | 7.7% | 7.4% | 14.3% | about 45 min | 〜240 MB |
Processing times are for our test computer's CPU (1 thread). Phones take longer. Whisper Medium and larger take several hours for 1 hour of audio on the CPU.
What the measurements show
- Larger models are not always more accurate. Whisper Large v3 is more accurate than Large v3 Turbo for Japanese and Korean, but makes more errors than Turbo for Spanish and Chinese, and takes about 1.4 times as long.
- For Japanese and Chinese, the 240 MB SenseVoice Small is as accurate as the 1.5 GB Whisper Large v3 Turbo or better (Japanese 8.6% vs 8.9%, Chinese 7.9% vs 9.9%).
- Whisper Medium is almost as accurate as Large v3 Turbo at less than half the download size (650 MB vs 1.5 GB). It takes about 1.2 times as long as Turbo.
- Qwen3-ASR 1.7B has the fewest errors for English, Spanish and Chinese.
- The fewest errors for Japanese come from Haas-ja v1.0 (5.2%), about 40% fewer than SenseVoice Small (8.6%). It also takes about a third of the time SenseVoice Small does.
- On the CPU, as on phones, Whisper Tiny's error rate rises sharply (from 44.4% to 70.5% for Japanese). Small barely changes.
How we measured
- The audio is 100 utterances per language, picked at even intervals from the test split of FLEURS (read-aloud speech), a multilingual speech recognition benchmark published by Google.
- We transcribed in the browser with the same software as this tool (transformers.js 4.2.0), the same model formats and the same processing (the initial text given to Whisper, and re-running sections that came back empty).
- We removed punctuation, spaces and full-width/half-width differences before comparing with the correct text. Chinese was converted to Simplified Chinese, and English went through the text normalization used in Whisper's evaluation.
- The test computer has an AMD Ryzen 7 8745H (Radeon 780M integrated graphics) and 32 GB of memory, and the browser was Chromium 153. Processing times do not include loading the model.
- Haas-ja v1.0 and Voxtral Mini 4B were measured for Japanese only, and SenseVoice Small for the 4 languages other than Spanish (only the languages each model can be selected for in this tool). Voxtral Mini 4B was measured the day before the other models and Haas-ja v1.0 on 3 November 2026, on the same computer and audio.
- Because Qwen3-ASR 1.7B uses a separate model to add timestamps, its processing time was measured by transcribing Japanese audio (108 seconds) in this tool, including that step.
Frequently asked questions
Are larger models more accurate?
Not always. For Whisper, accuracy improves with size from Tiny to Medium, but Large v3 made more errors than Large v3 Turbo for Spanish and Chinese. For Japanese and Chinese, the small SenseVoice Small is about as accurate as Turbo.
Which model should I choose on a phone?
For Japanese, Chinese, Korean and English, choose SenseVoice Small. It is also the default. For other languages, Whisper Small is the most accurate, but it is slow, so Whisper Base is also an option for long audio. Tiny makes many errors and is not recommended.
How is the default model chosen?
It is chosen automatically from the selected language and the device's capabilities. For Japanese, it is Haas-ja v1.0 on a computer with 4 GB of memory or more where WebGPU is available in Chrome, Edge or a similar browser (except Macs with Apple silicon). For Japanese elsewhere, such as on phones or in Safari and Firefox, and for Chinese and Korean, it is SenseVoice Small. For English and other languages, one of the Whisper models is chosen based on the GPU and memory. For English, devices that would get Whisper Base or smaller (such as phones and computers with integrated graphics) get SenseVoice Small instead.
Will the error rate be the same for meeting recordings?
No. These measurements use speech read aloud by one person, so every model makes more errors on noisy recordings or conversations between several people. Use them as a guide for comparing models.
Choose a model and transcribe
This tool is free and requires no sign-up. Audio files are processed in your browser and never leave your device.
Open the transcription tool