ebook2audiobook (E2A)

You can upload any voice audio to use as a reference for cloning. The ideal duration for a reference clip is between 1 and 5 million [units/samples, likely implying a specific length context provided in the full docs].

E2A automatically cleans the voice from background noise or music. The built-in cloned voice list is primarily in English; for other languages, you may need to contact the developers to have them officially added to the preset list.

You can use SML (Speech Markup Language) tags within your text to control pauses and voice switching during conversion:

• [break]: Inserts silence with a random duration between 0.3–0.6 sec.
• [pause]: Inserts silence with a random duration between 1.0–1.6 sec.
• [pause:N]: Inserts a fixed pause of N seconds (e.g., [pause:3] for a 3-second pause).
• [voice:/path/to/voice/file]...[/voice]: Switches the voice from the default or selected voice to the file specified in the path.

[pause:3]
[voice:/path/to/voice/file]This text uses a different voice.[/voice]

If you interrupt a training run using Ctrl+C, the training will stop, but the model will not be automatically packaged into a ready/ directory.

To use an interrupted model:

1. Monitor Quality: Run training with --sample-epoch-interval <N> to generate periodic audio samples in <run_dir>/epoch_samples/ so you can decide when to stop.
2. Stop: Press Ctrl+C when satisfied.
3. Export: Run the export_checkpoint.py script on your training run directory. This converts the checkpoint (e.g., to ONNX for Piper or optimized .pth for Coqui/XTTS) and creates a ready/ folder containing model.onnx and artifacts.json.

# Export a manually interrupted training run
python export_checkpoint.py /path/to/your/training_run/<timestamp>

You can insert SML (Speech Markup Language) tags into your eBook text to control pauses and voice switching during synthesis.

Available Tags

• [break]: Inserts a random silence between 0.3–0.6 sec.
• [pause]: Inserts a random silence between 1.0–1.6 sec.
• [pause:N]: Inserts a fixed pause of N seconds (e.g., [pause:3] for 3 seconds).
• [voice:/path/to/voice/file]...[/voice]: Switches the voice from the default or selected voice to the one specified in the path for the enclosed text.

Tip: For more advanced automation of SML insertion, see the E2A-SML repository.

[pause:3]
[voice:/path/to/voice/file]This text will be spoken with a different voice.[/voice]

You can use any voice recording in a supported audio format to clone a voice for text-to-speech. For best results, provide a recording between 1 and 5 minutes long. E2A includes built-in noise reduction that can clean up recordings even if they have background noise or music.

Note: The built-in list of cloned voices is primarily in English. For other languages, you may need to provide your own reference audio clip, especially when using custom XTTSv2 models.

You can insert SML (Speech Markup Language) tags into your text to control the pacing and voice of the generated audio:

• [break]: Inserts silence with a random duration between 0.3–0.6 sec.
• [pause]: Inserts silence with a random duration between 1.0–1.6 sec.
• [pause:N]: Inserts a fixed pause of N seconds (e.g., [pause:3]).
• [voice:/path/to/voice/file]...[/voice]: Switches the voice from the default/selected voice to the one specified in the path for the enclosed text.

E2A supports a wide variety of TTS engines and processing capabilities:

Supported TTS Engines:

• XTTSv2, Bark, Fairseq, VITS, Tacotron2, Tortoise, GlowTTS, YourTTS

Key Features:

• Multi-format conversion: Supports .epub, .mobi, .azw3, .fb2, .lrf, .rb, .snb, .tcr, .pdf, .txt, .rtf, .doc, .docx, .html, .odt, .azw, .tiff, .tif, .png, .jpg, .jpeg, .bmp, .zip.
• Text Box: Convert short snippets of text directly to audio.
• OCR Scanning: Processes files containing text as images.
• Voice Cloning: Optional feature using your own voice files.
• SML-tags: Provides precise control over interruptions, pauses, and voice changes.
• Custom Models: Support for custom trained models (XTTSv2, VITS, FAIRSEQ, PIPER, etc.).

You can upload any voice sample to be used for text-to-speech.

• Recommended duration: 1 to 5 minutes.
• Noise handling: E2A will automatically clean the audio, even if there is background noise or music.
• Note for XTTSv2: When using a specific XTTSv2 model, a reference audio clip is mandatory.

You can replace any voice sound using any supported audio format. The ideal duration for a reference clip is between 1 and 5 million (likely intended as 1 to 5 seconds or a specific sample length, though the text specifies '1 až 5 miliónov'). E2A automatically cleans the voice from background noise or music.

Note: The built-in list of cloned voices is primarily in English. To request official addition of voices in other languages, contact the developers.

You can upload audio clips to clone voices for use in supported audio formats.

• Ideal Duration: 1 to 5 minutes.
• Audio Quality: E2A automatically cleans up voices, even if the recording has loud background noise or music.
• Language Support: The built-in list of cloned voices is primarily English. To request official support for other languages, contact the developers.
• Requirement: For custom XTTSv2 models, a reference audio clip of the voice is mandatory.