A clear next step.

Your first audio

Start with one short paragraph. Choose a model, an output language and a compatible voice, then listen to the result before working on a longer project.

Model downloads need a network connection. Once prepared, local models process your script and audio on your Mac. External Text AI providers are a separate, optional connection.

System requirements

Sonano is a native workspace for Apple Silicon Macs running macOS 26 or later. Local models use your Mac’s memory and storage; larger models require more resources.

Different models support different languages, voice types and controls. A downloaded model is only ready when its runtime dependencies are also available.

Find your way around

The workspace keeps creation, reusable assets and results together. Moving between pages does not create a separate generation system.

Use Connect Agent for host setup and local capabilities. Use Settings for appearance, playback, the local engine, Text AI, storage and diagnostics.

Choose a workflow

For automation, start with Agent setup. The App and Agent share the same models, jobs and artifacts.

Choose a model

Start with the task, target language and voice source. Model Library shows the current variants, dependencies and readiness. Similar model names do not imply the same inputs or controls.

• Read with built-in voices: Kokoro has language-matched built-in voices. Qwen3 CustomVoice has 9 fixed speakers; choose 1.7B for style instructions.
• Reuse a reference voice or cast dialogue: choose a reference-capable route such as Qwen3 TTS Base, VoxCPM 2, multilingual Chatterbox or OmniVoice. Check model, language and asset compatibility. Character organization belongs to Sonano.
• English reference speech: Chatterbox Turbo is English only. Choose the multilingual route for other languages.
• Design a new voice: use Qwen3 VoiceDesign or VoxCPM 2, then save a reference for reuse. OmniVoice also offers design, trained mainly on Chinese and English design data. Its 600+ language coverage does not imply equally reliable design in every language.
• Transcription and subtitles: compare Qwen3-ASR and Whisper. Sonano exposes ten Qwen3-ASR language choices and finer alignment for Chinese, English, Japanese and Korean. Whisper provides wider language choices and segment timing. Alignment matches existing text to audio; it is not automatic text correction.
• Change an existing recording’s voice: use Chatterbox VC, API model ID chatterbox-vc. Provide source audio and a managed target with reference audio. Built-in presets are not conversion targets.

Test a short passage for language, voice and local resources before a longer project. See all 12 entries, integration boundaries and sources in the model guide. Discover actual API IDs through the manifest and GET /v1/models; VoxCPM 2 uses voxcpm.

Create speech, scripts and audiobooks

Speech has three modes. They share the model and voice settings, generation history, player and export system.

Normal speech

Write or import a script, choose your model, output language and voice, then generate. Longer text can be split into segments. The available style instructions and inline tags depend on the model; only use the controls shown for your selection.

Multi-Role scripts

Organize the script into speaker blocks. Give each speaker a suitable voice, then edit or generate a single block before committing to the full script. Script analysis can suggest speakers and blocks when Text AI is configured; review its suggestions before generating.

For example, a narrator can set the scene, a bookseller can offer a clue and a visitor can respond. Distinct voices make the speaker mapping clear. Listen to individual blocks, then generate the complete script.

Audiobook projects

Import EPUB, TXT or Markdown and review the proposed chapters before confirming them. Edit chapter titles, text and order. The current audiobook workflow uses one project-level model, language and voice.

• Generate a selected chapter or missing chapters.
• Review completed chapters in the player.
• Merge completed chapters into M4B, M4A, MP3 or WAV.
• M4B includes basic chapter metadata; an audiobook manifest records the project’s output structure.

App and HTTP workflows support import review, persistent chapters, generation runs and exports. The current MCP tool set does not expose audiobook tools; HTTP support does not automatically add a tool to your host. See API quick reference. Current text import does not provide PDF or OCR. Preview a chapter first and check the source text before generating a long book.

Keep and reuse a result

Completed generations enter History. Play, reuse or export them without generating again. When a job fails, inspect its error and model status before retrying; a failed request is not a completed audio result.

Transcribe and review media

Import audio or video, choose a model and language, then transcribe. Video input is processed through its audio track; a video with no usable audio cannot produce a transcript.

Review text and timing

Check the transcript against the media. Use the segment or unit timeline to locate mistakes and inspect subtitle boundaries. Search the transcript for a word, correct its text and preview the relevant media before applying changes.

Review suggested edits

Silence review identifies pause intervals. Text review can suggest changes to filler words, repetition or restarts when Text AI is configured. Review and accept the suggestions you want; these are separate steps and should not be treated as automatic audio-quality guarantees.

Export the right timeline

When you edit media timing, export the matching edited subtitles. Source-timeline subtitles and edited-media subtitles serve different purposes.

Convert a recording’s voice

Voice Conversion takes an existing recording and a target voice. It aims to preserve the source content and delivery while changing its timbre. It is different from typing a new script and generating speech.

1. Import the source recording and listen to check it.
2. Choose a managed factory, designed or cloned voice with reference audio.
3. Check that the conversion model is ready, then run the job.
4. Compare the result with the source and export the version you want.

Model presets do not necessarily include reusable reference audio and cannot be used as conversion targets. Create or choose a saved voice asset instead.

Start with clean, clear, single-speaker audio. Voice conversion does not guarantee the removal of noise, overlapping speech or distortion. Only use voices and recordings you are authorized to use.

Build a reusable voice library

Voices can be reused across Speech, compatible conversion workflows and Agent calls. Search by name, language, gender or tags, preview a reference sample and keep frequently used voices in your favorites.

Design or clone

Choose New Voice, then Design Voice or Clone Voice. For design, describe the voice, generate candidates, listen and save the one you want. For cloning, provide an authorized reference recording and its text, then review the saved asset. Follow the selected model’s duration and input guidance rather than a universal recording length.

Choose a model and output language before browsing voices in Speech. The App filters the list for compatible choices. A voice asset’s reference sample is not a guarantee of every model’s performance.

Export files

Sonano keeps results as managed artifacts. Export a copy when you are ready to use the result in an editor or share it.

Free and Pro use the same current models, generation quality and export formats. Pro changes local usage and saved-voice limits, not the quality of a file you already made.

Deleting a managed history result does not remove copies you exported elsewhere. Keep a separate backup of finished work.

Connect an Agent

Use Connect Agent in the App: choose your Agent, connect Sonano, then test inside the Agent. A saved configuration is preparation, not proof that a tool call succeeded.

MCP, the Skill and an optional voice backend

Cursor, Claude Code, Codex, OpenCode, OpenClaw and Hermes have host-specific MCP setup. Pi currently has Skill and manual-connection guidance. Available actions depend on the host version, configuration location and permissions; follow the current App. Use component management actions to remove a setup or deliberately reinstall it.

What the tools can do

More connection options and tools

• Service details and diagnostics: inspect the current address, version and health. The default is http://127.0.0.1:19280; update host configuration after a port change.
• Manual setup and shared Skill: find the shared directory, Skill path, MCP configuration and connection message for another Agent. Copy commands and launcher paths from your actual App.
• Developer / API: inspect capabilities, examples and job events. See API quick reference for HTTP endpoints and artifact retrieval.

External media and local authentication

File access depends on host and client permissions. Do not assume that a sandboxed MCP process can read any file_path. The installed Skill’s host-side media upload flow can submit bytes as a transcription or conversion job, then query its job and artifacts. It requires terminal, file and loopback access in the host; it does not grant broader sandbox permissions.

curl -i http://127.0.0.1:19280/health

Only GET /health and HEAD /health are public. Every other API, including discovery, models and voices, requires same-user local authentication. The App, MCP and official Skill read credentials automatically; no manual pairing is required. X-Sonano-Caller is attribution, not authentication. Never share a local token in a website, screenshot or public log.

Discover capabilities first, then check model, language and voice compatibility. Text-processing tools also need Text AI configured. See the Agent overview and model guide.

API quick reference

The default HTTP address is 127.0.0.1:19280; use the App’s actual port. MCP exposes a tool set over the underlying HTTP API. Their capability surfaces are not identical.

Use the current manifest and App examples for parameters, errors and delivery. MCP does not launch the App, download models automatically or resume every ordinary job after an engine restart. Query the original job after reconnecting before deciding to retry.

Except GET/HEAD health checks, these APIs require same-user local authentication. X-Sonano-Caller alone does not authenticate a request. See Agent setup for connection and external-media upload guidance.

Track jobs, results and requests

Jobs & Progress

Downloads, generation, transcription and conversion appear in one task list. Inspect queued, running, completed, canceled or failed work and read the actual error before retrying.

History

Completed work is retained with the information needed to preview, reuse and export it. Find a result by its text, file, model, voice or task context.

Request Audit

External calls can leave local records with caller, capability, request identifier, endpoint and outcome. These help connect an Agent request to the resulting job or artifact.

Clearing completed task entries is different from deleting history results and artifacts. Read each cleanup action’s scope before confirming it; exported copies remain separate.

Download and manage models

Open Model Library to see model families, variants, languages, capabilities, dependencies and download status. Choose a variant that fits your Mac and task.

• Not downloaded: the weights are not available in the managed cache.
• Missing dependencies: required local-engine components are not ready.
• Ready: the selected model can be used for supported tasks.
• Current selection: a preference, which does not by itself mean the weights are downloaded.

Downloads appear in Jobs & Progress. Current OmniVoice and VoxCPM2 use the shared MLX runtime; do not follow older instructions for separate official-model runtimes.

Use the App’s model and storage management tools for cleanup. Download size, model parameter count and available system memory are different measurements.

Set up your workspace

Microphone permission is needed when recording a voice-clone sample or re-recording a media passage in the App. Settings are saved locally. Changing the interface language does not translate a script, recording or model identifier.

Local data and upgrades

Your models, voices, history and results are stored separately from the App bundle. Replacing the App normally keeps this managed data; deleting Sonano.app does not automatically erase it.

• Open Settings → Storage to inspect categories and their locations.
• Export finished work to a location you control and maintain your own backups.
• Use managed cleanup controls instead of deleting unknown directories while jobs are active.
• The website edition commonly stores logs at ~/Library/Application Support/Sonano/logs/. The Mac App Store edition uses its sandbox container; custom configuration can also change locations. Use Settings → Storage for the actual paths.

Imported media, model caches, temporary inputs and generated artifacts have different purposes. A file that is not obvious in History should not be assumed safe to delete.

Local processing and external services

Downloaded local models process scripts and audio on your Mac. This does not mean every feature works without a network connection.

• Model downloads contact their sources.
• Purchases, activation, recovery and periodic license checks use online services.
• Optional external Text AI providers receive the content you ask them to process.
• Local Agent and API calls use the engine’s authentication and managed job system.

Review provider settings before sending sensitive material. Voice cloning requires appropriate permission; Pro does not grant rights to another person’s voice or override model licenses.

Read the local-data overview and privacy policy for more detail.

Updates, licenses and purchase channels

Start with Free and upgrade to Pro when you need more usage. A website license covers Sonano 1.x and up to two active Macs. It can be used offline for up to 14 days after a successful online verification.

For a website purchase, enter the license key in the App. Recover keys and deactivate old Macs by verifying the purchase email on the website. The App does not require email sign-in.

Mac App Store purchases are restored in that edition through Apple. Website and Mac App Store purchases do not automatically unlock each other. Refund procedures follow the purchase channel.

Check the current version on the download page and changes in the release notes. For activation, recovery or refunds, use purchase help.

Find the cause of a problem

Keep the job and request identifiers with the relevant log excerpt. Do not repeatedly purchase, redownload or regenerate to diagnose an uncertain status.

Common questions

Prepare a useful support request

Send enough context to connect the issue to a model, job and result. A short, relevant log excerpt is more useful than an unfiltered data folder.

Remove license keys, verification codes, tokens, API keys and unrelated private content before sending a report.

Contact support

Write to [email protected] with the details from Prepare a useful support request.

For payments, recovery, changing Macs or refunds, start with purchase help and mention your purchase channel. Do not include your license key or a verification code.

For an App or Agent issue, include the model, workflow, approximate time and relevant task identifiers so the local sequence can be understood.