Skip to content

Generate Voice (TTS)

Synthesise subtitle files (with timing) or arbitrary text into MP3 / WAV audio. Five TTS backends: Edge TTS (free), ElevenLabs (high quality), Google Cloud TTS, Gemini TTS (free tier), and Piper TTS (offline).

What you need

  • FFmpeg on PATH — see FFmpeg setup.
  • A TTS backend, one of:
    • Edge TTS — free, no key, default. Uses Microsoft Edge's cloud voices.
    • ElevenLabs — paid, highest quality. See ElevenLabs setup.
    • Google Cloud TTS — paid, very good. See Google Cloud setup.
    • Gemini TTS — free tier, natural prebuilt voices. Reuses your existing Gemini API key from the LLM tab — no extra setup.
    • Piper TTS — fully offline neural TTS. No API key, no network calls — voices are ~25–60 MB ONNX files downloaded once via Settings → Voice → Piper TTS → Download voices now. 32 of the app's 45 languages have a Piper voice today; languages without Piper coverage silently fall back to Edge TTS at synthesis time.

Walkthrough

  1. Click Generate Voice in the sidebar.
  2. Drop one or more .srt / .vtt / .ass / .ssa subtitle files.
  3. Pick the Language (auto-detected from the subtitle filename when possible — e.g. _translated_en_fr.srt is detected as French).
  4. Pick Voice gender — Female or Male.
  5. Pick Output format — .mp3 (default) or .wav.
  6. Click Generate (or Ctrl+Enter).
  7. Open the row when done — it plays in your default audio app.

Output

You get a single audio file with the voice tracks placed at each subtitle's timestamp. Silent gaps fill the time between cues so the audio stays in sync with the original timing.

Picking a TTS backend

Backend Cost Voices Notes
Edge TTS Free Hundreds, all major languages Default. No setup.
ElevenLabs Paid (~$5/mo entry tier) Premium neural voices, voice cloning Highest quality. Voice ID is set in Settings → Service.
Google Cloud TTS Paid (~$4/M chars; 1 M free / month) WaveNet / Studio voices in 50+ languages Strong WaveNet voices for European languages. By default the server picks a voice based on language + gender.
Gemini TTS Free tier (Developer API quotas apply) Natural prebuilt voices in 24+ languages — Kore (female default) / Puck (male default) Reuses your Gemini API key from the LLM tab. Per-call output capped at ~30 s; long text chunks at sentence boundaries automatically.
Piper TTS Free, offline Neural voices in 32 of the app's 45 languages No key, no network. Per-language voice downloaded on demand from Settings → Voice → Piper TTS → Download voices now (~25–60 MB each). Pre-flight catches a missing voice before work starts.

Switch in Settings → Voice → TTS method.

Piper TTS specifics

Piper is the only fully-offline TTS backend in the app. A few things to know:

  • Voice library dialogue — open via Settings → Voice → Piper TTS → Download voices now. Each language row shows a Female voice and / or Male voice download button (some languages are single-gender). Voices come from the rhasspy/piper-voices HuggingFace catalogue.
  • Coverage — 32 of the app's 45 languages have a Piper voice. The 13 without coverage (Belarusian, Bengali, Chinese (Traditional), Croatian, Estonian, Hebrew, Japanese, Khmer, Korean, Lithuanian, Malay, Mongolian, Thai) silently fall back to Edge TTS at synthesis time so synthesis never hard-fails on a missing voice.
  • Gender resolution — when you pick Female, the engine first tries the female voice for that language; if only a male voice exists, it uses that instead (and vice versa). Logged at INFO level.
  • Pre-flight gate — before a Voice run starts, the page checks that the per-language Piper voice is on disk. If missing, you get a modal dialogue with an Open Settings button that takes you straight to the voice library so you can download it without losing your queue.

Gemini TTS specifics

Gemini TTS uses gemini-2.5-flash-preview-tts via the Developer API. A few things to know:

  • Voice selection is by gender today — Female maps to Kore, Male to Puck. Both are clear, neutral voices that work across languages without sounding too character-y.
  • Output length cap — each Gemini API call returns at most ~30 s of speech. The app chunks input text under _GEMINI_TTS_MAX_BYTES (~2000 bytes ≈ 30 s) at sentence boundaries, then concatenates the chunks via FFmpeg. You won't hit truncation on normal subtitle text.
  • Audio format — Gemini emits raw PCM at 24 kHz mono s16le; the app transcodes per-chunk to MP3 (or WAV if you picked it) so the final file matches your selected output format.
  • Vertex AI is not yet supported for TTS — even if your LLM tab is configured for Vertex, Gemini TTS still needs a Developer API key. The app raises AUTH_ERROR upfront if missing.

ElevenLabs models

Three models are exposed:

Model Latency Quality Use for
eleven_multilingual_v2 (default) Medium High General TTS
eleven_v3 Medium Highest Studio / production
eleven_flash_v2_5 Low Good Real-time / Live mode

Configure in Settings → Voice → ElevenLabs model.

Tips

Re-generate

Right-click a row → Re-generate to swap voice gender / TTS method / format without re-running translation.

Pre-flight checks

The page validates ElevenLabs API key (when selected) and FFmpeg availability before starting. You'll see a friendly dialogue if anything is missing.

Stop is atomic

Hit Stop during synthesis and you won't get a half-written MP3 in the output directory — the file is written to a temp location first, then moved into place only on success.

Shortcuts

Shortcut Action
Ctrl+Enter Generate
Ctrl+O Browse
Ctrl+F Focus history search