Start with the deliverable
An interview can become an edited article or a sequence of captions. The article needs accurate, editable text. Captions also need useful timing and readable segmentation. Decide what you will deliver before choosing a model.
Sonano offers Qwen3-ASR and Whisper. Compare them on the same original audio with the same language and range. A cleaned speech track and a different recording with music and overlapping voices are not a controlled comparison.
Check the app’s language choices
Qwen3-ASR upstream offers 0.6B and 1.7B models and lists 30 languages plus 22 Chinese dialects. Sonano currently exposes ten language choices for this route. Its Whisper options include Tiny, Base, Small, Medium, Large-v3 Turbo and Large-v3.
Check that your target language is available in the current selector. Test dialects, mixed-language terminology and accents on real excerpts. This guide does not combine scores from different upstream datasets into a single ranking.
Recognition and alignment do different jobs
The Whisper route provides segment timing. The Qwen3-ASR workflow can use Qwen3-ForcedAligner for finer word or character timing in supported languages. Sonano currently exposes that fine-alignment path for Chinese, English, Japanese and Korean.
Forced alignment uses audio and known text to estimate where the words occur. It does not automatically correct an inaccurate transcript. Check the words first, then review boundaries around pauses, fast speech and speaker changes.
Keep your own comparison record
Use a short recording with clear speech, specialist terms and one difficult passage. Run two candidates and note the following. These are suggested checks, not measured results from this site.
| Record | Check |
|---|---|
| Recognition errors | Omissions, repetition, names and numbers against playback |
| Subtitle timing | Speech starts, pauses and sentence changes |
| Waiting and editing | Processing time and manual correction time separately |
| Resource use | Whether the same Mac stays usable for other work |
A model that handles one difficult passage particularly well may become a useful fallback. Choose the everyday model around the recordings you process most often.
Finish with the right export
After review, export TXT for a transcript, SRT or VTT for timed subtitles, or JSON for structured data. If you removed pauses or sections from the media, use subtitles that match the edited timeline.
Continue with the local SRT workflow or the model selection documentation.
Sources & further reading
- Qwen3-ASR · Recognition and alignment
- OpenAI Whisper · Model family
- Sonano · Current transcription options
Sources checked on November 3, 2026. Available app options depend on the current version.