Local models · Choose by task
Speech, transcription and voice creation call for different tools. Choose, download and manage them in one library, then get back to your work.
Start with the task
Start with a task, then consider language, voice source and memory. Free and Pro use the same models and generation quality.
These descriptions follow the models and variants currently integrated in Sonano. Upstream capabilities are not automatically App features. Character organization, candidate management and subtitle export are Sonano workflows.
An 82M-parameter speech model with built-in voices for read-throughs and everyday narration.
English (US/UK), Japanese, Mandarin, Spanish, French, Hindi, Italian and Brazilian Portuguese. Match voices to their language; voice availability and results vary. This route does not clone a reference recording.
Model reference ↗Reuse a reference voice for narration. Sonano organizes separate voices and lines into multi-role dialogue.
Base uses reference audio through Sonano’s factory, designed or authorized cloned assets, with streaming preview. Languages: Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian. Current Base variants are all 0.6B; 4bit/8bit/bf16 are precision options, not different parameter counts. Base has no natural-language style instruction control.
Model reference ↗Choose from 9 built-in speakers. The 1.7B version also supports natural-language style instructions.
Sonano offers 0.6B and 1.7B variants. Choose 1.7B when you want to guide expression with style instructions. This route does not accept your own reference audio; it is distinct from cloning and voice design.
Model reference ↗Describe a voice and generate sample candidates, then save your choice in Sonano’s voice library.
The 1.7B VoiceDesign model supports the Qwen3 set of ten output languages. Design creates a new voice; Base later reuses the saved reference. Saving, comparing candidates and organizing characters are Sonano workflows.
Model reference ↗One model for text-to-speech, description-based voice design and reference-voice cloning with style guidance.
VoxCPM2 covers 30 languages including Chinese, English, Japanese, Korean, French and German. Describe a new voice without a reference, or guide expression when using a reference. Compare design or cloning results with a short passage first.
Model reference ↗An English speech generation and voice-cloning route for narration and dialogue.
Turbo targets lower-compute English generation, not multilingual speech. Sonano uses managed reference voices; speed depends on the variant and your Mac. General natural-language style instructions are not currently accepted.
Model reference ↗The multilingual Chatterbox route integrated in Sonano generates new text using a reference voice.
23 languages including Chinese, English, Japanese and Korean, for multilingual reading with a reference voice. For English-only work, you can also compare a Turbo preview. This route does not offer voice design or general natural-language style instructions.
Model reference ↗A reference-voice cloning route for narration and localization across a wide range of target languages.
Upstream reports 600+ languages. Voice design is also available, but its design training uses Chinese and English data. Cloning is the main training task; preview low-resource languages, cross-language accents and designed voices before use. Coverage does not imply equal quality.
Model reference ↗Recognize speech in audio or video, then review transcripts and prepare subtitles in Sonano.
Upstream supports 30 languages and 22 Chinese dialects. Sonano currently exposes ten language choices: Chinese, English, Japanese, Korean, French, German, Spanish, Italian, Portuguese and Russian. For word or character timing, pair it with ForcedAligner in a supported language.
Model reference ↗Choose a local transcription model size to suit your language, recognition needs and Mac resources.
Sonano offers Tiny, Base, Small, Medium, Large-v3 Turbo and Large-v3. In Sonano, use it for transcription and segment timing. Compare variants with the same recording to find a fit for your language and Mac resources. Results vary by language and recording.
Model reference ↗Use audio and an existing transcript to estimate word or character start/end times for review and subtitles.
This is not a standalone recognizer or automatic text corrector. Upstream supports 11 languages; Sonano currently exposes finer alignment for Chinese, English, Japanese and Korean in its Qwen3-ASR timestamp workflow.
Model reference ↗Use source audio and a target voice reference to change vocal timbre, without generating a new reading from text.
Choose an existing recording and a target voice with reference audio. A regular built-in voice may not be a valid target. Preview the converted timbre, timing and background. To generate new content from text, choose a speech generation model.
Model reference ↗Prepare once. Create locally.
Download weights, choose variants and check dependencies in the app. These explain the states; they are not live readings from your Mac.
The app identifies the model and variant you need. Downloads appear in Jobs & Progress, even when you switch workspaces.
Weights, runtime dependencies and your current selection are shown separately. Selecting a model does not mean it is downloaded or ready.
Check that your language and voice are compatible, then preview a short passage. Speed and available variants depend on the model and your Mac.
Model downloads need a connection; downloaded local models can run offline. Model licenses and rights to voice materials still apply.
From a line of narration to dialogue and a whole audiobook.