Another TTS

Easy Workflow for creating Training Data for TTS.

Installation

Docker-Way

git clone https://github.com/ItsJamin/another-tts
cd another-tts
docker compose up

Python-Way

• Create .venv (Python 3.13.7) python -m venv .venv

• Activate Virtual Environment Linux: source .venv/bin/activate Windows: .venv\Scripts\activate

• Install Libraries: pip install -r requirements.txt

• Install ffmpeg if not in system (needs to be available on cmd with ffmpeg)

• Create in the env/ directory a .env-File with the following content:

CURRENT_DATASET = "<your_dataset_name>"
CURRENT_LANGUAGE = "de" # currently only german sentences

Change the name of your dataset to what your dataset should be called.

• Start Server with python app.py and visit localhost:5067

Data

• data/sentences/CURRENT_LANGUAGE/ - Textfiles of what to say
• data/datasets/CURRENT_DATASET/ - Place where recordings are saved.
• data/datasets/CURRENT_DATASET/metadata.csv - Which audiofiles contain what text. (see example Dataset)

TODO

• GUI for easily recording Data

• GUI for Overview over recorded Data

• Docker Implementation

• Filter out or show already recorded lines

• Metadata for voice (mood, whisper, etc.)?

---

Standards

What criteria should lines follow?

• Transcript text must match the spoken audio verbatim
• No paraphrasing, omissions, or additions
• All text must be fully normalized (numbers, dates, abbreviations, special symbols written out)
• DO use points, question marks, commatas and apostrophs.
• No inconsistent or ambiguous punctuation
• One sentence per line
• Neutral, non-acted language (does not mean boring)
• Lines should vary in length (short, medium, long) to ensure phonetic coverage
• Questions, statements, numbers, dates, and abbreviations must be represented
• The same text normalization rules must be applied consistently across the entire dataset

Examples:

• I have 42 apples. -> I have forty two apples.
• The contract was signed in 2025. -> The contract was signed in twenty twenty five.
• Dr. Smith will arrive at 9 a.m. -> Doctor Smith will arrive at nine a m.
• The battery is at 80%. -> The battery is at eighty percent.
• "Wait, how did you 2 meet again?" -> Wait, how did you two meet again?

---

What criteria should recordings follow?

• WAV format, PCM (uncompressed), mono
• Consistent sample rate across all files
• Recommended: 44.1 kHz, 16-bit
• One sentence per file
• File duration ideally between 1.5 and 8 seconds
• No leading or trailing silence greater than ~200 ms
• No clipping, distortion, or background noise
• Stable loudness across all recordings
• Target: −20 to −16 LUFS, peak below −1 dBFS
• Recorded in the same environment with the same microphone and setup
• Fixed microphone distance and speaking posture
• Consistent speaking style, pace, and energy level
• No whispering, shouting, or expressive acting
• Speaker must be healthy and vocally consistent across sessions
• File names must be deterministic, sortable, and space-free (e.g. speaker_0001.wav)
• Stereo files, variable sample rates, and inconsistent bit depth are not allowed