Skip to main content
ChatbotAI-Free supports two TTS engines: built-in Kokoro voices for English and Spanish, and optional Sherpa-ONNX voice packs for additional languages.

​
Built-in Kokoro Voices

Kokoro TTS v1.0 provides 54 high-quality voices out of the box, including both English and Spanish speakers.

​
Voice Files

Kokoro requires two files in the voices/kokoro-v1.0/ directory: 
voices/
└── kokoro-v1.0/
    ├── kokoro-v1.0.onnx    ← ~300 MB neural TTS model
    └── voices-v1.0.bin     ← ~27 MB  (54 English + Spanish voices)
Download these files from the kokoro-onnx releases page. They’re too large to include in the repository.

​
Voice Loading

The Kokoro wrapper loads voices from the binary file using NumPy: 
# From kokoro_wrapper.py:28-30
data = np.load(voices_path, allow_pickle=False)
self.voices = {name: data[name].astype(np.float32) for name in data.files}

​
Voice Naming Convention

Kokoro voices follow a specific naming pattern:
  • Format: {age}{gender}_{name}
  • Age codes: a (adult), y (young), o (old)
  • Gender codes: f (female), m (male)

​
Examples

Voice NameDescription
af_bellaAdult female - Bella (default English voice)
af_sarahAdult female - Sarah
am_adamAdult male - Adam
ef_doraSpanish adult female - Dora (default Spanish voice)
yf_emmaYoung female - Emma
Kokoro voices never contain hyphens in their names. This is used by the TTS manager to distinguish them from Sherpa-ONNX voices (see tts_manager.py:22).

​
Adding Sherpa-ONNX Voice Packs

Expand language support by adding Piper-compatible Sherpa-ONNX voice packs for French, Italian, German, Portuguese, and more.
1

Install Sherpa-ONNX

First, install the sherpa-onnx Python package:
pip install sherpa-onnx
2

Download a voice pack

Browse available voices at huggingface.co/csukuangfj.Download these files from the voice pack repository:
  • The .onnx model file (e.g., es_AR-daniela-high.onnx)
  • tokens.txt
  • espeak-ng-data/ directory
3

Place in voices/ folder

Create a folder inside voices/ with the voice pack files:
voices/
├── kokoro-v1.0/
└── vits-piper-es_AR-daniela-high/  ← your new voice
    ├── es_AR-daniela-high.onnx
    ├── tokens.txt
    └── espeak-ng-data/
4

Restart the app

The voice scanner will detect the new pack and prompt you to classify its language.
Sherpa voice folder names must contain hyphens to be distinguished from Kokoro voices. The app uses this pattern to route synthesis to the correct engine.

​
Voice Scanner Auto-Detection

On startup, the voice scanner automatically detects installed voices:

​
Detection Criteria

The scanner identifies Sherpa-ONNX voice packs by checking for:
  1. At least one .onnx file in the folder
  2. An espeak-ng-data/ subdirectory
From voice_scanner.py:55-67: 
def is_sherpa_voice_folder(folder_path: Path) -> bool:
    if not folder_path.is_dir():
        return False
    has_onnx = any(folder_path.glob("*.onnx"))
    has_espeak = (folder_path / "espeak-ng-data").is_dir()
    return has_onnx and has_espeak

​
Voice Configuration Storage

Classifications are saved to voices/voices_config.json so you’re only prompted once per voice pack: 
{
  "vits-piper-es_AR-daniela-high": "spanish",
  "vits-piper-fr_FR-siwis-medium": "french"
}

​
Language Classification for New Voices

When a new Sherpa voice is detected, the app shows a one-time dialog asking you to assign it to a language category.
The TTS manager needs to know which language setting should activate each voice. This allows the app to automatically switch voices when you change the language in Settings.
Edit voices/voices_config.json manually and change the language assignment, then restart the app.

​
Classification Flow

From voice_scanner.py:130-142: 
if is_sherpa_voice_folder(entry):
    onnx_files = list(entry.glob("*.onnx"))
    onnx_path = str(onnx_files[0]) if onnx_files else None
    cached_lang = config.get(name)  # "english" / "spanish" / None
    result["sherpa_voices"].append({
        "folder": name,
        "path": str(entry),
        "onnx": onnx_path,
        "language": cached_lang,
        "is_new": cached_lang is None,
    })

​
Voice Selection

Choose your active voice from the top dropdown menu in the main interface. The list includes:
  • All 54 built-in Kokoro voices
  • Any installed Sherpa-ONNX voice packs (classified by language)
Voice preferences are saved per-language:
  • english_voice: Default voice for English (default: "af_bella")
  • spanish_voice: Default voice for Spanish (default: "ef_dora")
See preferences.py:34-35 for default voice assignments.

Build docs developers (and LLMs) love

Get started for freeTalk to us