Skip to content

Balaka v0.3.0 — safer local TTS

Latest
Latest

Choose a tag to compare

View all tags
@stremovskyy stremovskyy released this 27 Oct 09:19
v0.3.0
f342c67
This commit was signed with the committer’s verified signature.
stremovskyy Anton Stremovskyy
GPG key ID: B9B858F8FAC98F4F
Verified
Learn about vigilant mode.

Balaka runs speech synthesis on your computer. This release updates the OmniVoice runtime, tightens local-server defaults, and makes the browser interface easier to use and explore.

Highlights

  • OmniVoice 0.2.1: Voice design and voice cloning now use the current runtime, with verified 24 kHz WAV output from both modes. The same k2-fsa/OmniVoice model is used, but the new inference implementation can change how a voice sounds compared with v0.2.0.
  • Local by default: The server binds to 127.0.0.1:8000. BALAKA_HOST and BALAKA_PORT allow an explicit override.
  • Safer requests: Reference recordings are limited to 20 MiB (413 when exceeded). One synthesis runs at a time per process; another request gets a retryable 503 response.
  • More control in the UI: Advanced settings now expose silence padding and fade duration. A new Text tips dialog explains supported text controls and their limits, while Show API request displays the current form fields and a copyable cURL command.
  • Reproducible setup: Web and runtime dependencies are pinned. CI uses current GitHub Actions versions and runs the API test suite on Python 3.13.

Upgrade notes

Use Python 3.13 and recreate or update your virtual environment using the setup instructions. The default model is unchanged. If a familiar voice sounds different, compare the same text, voice settings, and reference recording; voice-design output can vary between generations.

The server is accessible from this machine by default. To connect from another device, set BALAKA_HOST explicitly and account for your network's access controls.

Verification

  • 18 API and service tests passed locally.
  • Python compilation, JavaScript syntax, dependency consistency, and diff checks passed.
  • Real Ukrainian voice-design and voice-clone generation produced playable 24 kHz mono WAV files on Apple Silicon.

Listen to v0.3.0

Links

Assets 4
Loading