py-speech-gen 0.1.0

A Python library for generating synthetic speech datasets using TTS providers.

py-speech-gen

A Python library for generating synthetic speech datasets using TTS providers. Supports ElevenLabs and Piper TTS out of the box, with an extensible provider system for adding custom backends.

Features

• Multi-provider support — ElevenLabs, Piper TTS, or your own custom provider
• Text preprocessing — cleaning, normalization, number-to-words, sentence segmentation
• Parameter randomization — per-sample variation for voice diversity
• Background noise injection — 8 synthetic noise types (white, pink, brown, traffic, cafe, home, crowd, mic)
• Flexible output formats — WAV, MP3, FLAC at configurable sample rates
• Reproducible generation — export/load configs for deterministic datasets
• Export options — JSON, CSV, pandas DataFrame

Installation

Basic install (core features only)

pip install py-speech-gen

The base installation includes text processing, dataset management, randomization, and noise mixing.

Install with TTS providers

# Piper TTS (local, offline)
pip install "py-speech-gen[piper]"

# ElevenLabs (cloud API)
pip install "py-speech-gen[elevenlabs]"

# All providers
pip install "py-speech-gen[all]"

Requirements

• Python 3.11+
• For Piper TTS: ONNX Runtime (GPU or CPU variant)
• For ElevenLabs: valid API key

Quick Start

import os
from pathlib import Path
from py_speech_gen import (
    ElevenLabsProvider,
    PiperProvider,
    DatasetGenerator,
    TextProcessor,
)

# Setup providers
elevenlabs = ElevenLabsProvider(
    api_key=os.getenv("ELEVENLABS_API_KEY"),
    voice_ids=["voice_id_1", "voice_id_2"],
    model="eleven_flash_v2_5",
)

piper = PiperProvider(
    models_path="./models",
    voice_ids=["en_US-lessac-medium"],
)

# Generate dataset
generator = DatasetGenerator(
    providers=[elevenlabs, piper],
    output_dir="./output",
)

dataset = generator.generate_dataset(
    texts=["Hello world.", "This is a test."],
    dataset_name="my_speech_dataset",
    process_texts=True,
    texts_lang="en",
)

# Export results
dataset.save()        # JSON
dataset.save_csv()    # CSV
print(f"Generated {len(dataset)} samples")

Components

Component	Description
Providers	TTS backends (ElevenLabsProvider, PiperProvider) implementing BaseProvider
DatasetGenerator	Orchestrator that manages generation across multiple providers
Dataset	Data model with save/load/export (JSON, CSV, pandas)
TextProcessor	Text cleaning, normalization, number-to-words, sentence segmentation
Randomizer	Per-sample parameter randomization for voice diversity
NoiseMixer	Background noise injection for realistic conditions

Randomizer

Adds random variation to provider parameters per sample for dataset diversity.

from py_speech_gen import Randomizer

# Use a preset: "subtle", "moderate", "extreme"
randomizer = Randomizer.preset("moderate", seed=42)

generator = DatasetGenerator(
    providers=[piper, elevenlabs],
    output_dir="./output",
    randomizer=randomizer,
)

Custom Randomizer Config

randomizer = Randomizer(
    config={
        "global": {"speed": (0.8, 1.2)},
        "elevenlabs": {
            "stability": (0.3, 0.7),
            "similarity_boost": (0.5, 0.9),
            "style": (0.0, 0.3),
            "use_speaker_boost": [True, False],
        },
        "piper": {
            "length_scale": (0.8, 1.3),
            "noise_scale": (0.5, 0.8),
            "noise_w_scale": (0.6, 1.0),
        },
    },
    seed=42,
)

# Export/load for reproducibility
randomizer.export("randomizer_config.json")
r = Randomizer.load("randomizer_config.json")

NoiseMixer

Adds background noise to generated audio for realistic conditions.

from py_speech_gen import NoiseMixer

mixer = NoiseMixer(
    noise_types=["white", "traffic", "cafe"],  # or "all", "synthetic"
    snr_db=20,
    random_snr=True,
    snr_range=(15, 30),
    skip_prob=0.2,  # 20% of samples get no noise
    seed=42,
)

generator = DatasetGenerator(
    providers=[piper, elevenlabs],
    output_dir="./output",
    noise_mixer=mixer,
)

Noise Types

Type	Source	Description
white	Generated	White noise
pink	Generated	Pink noise (1/f)
brown	Generated	Brown noise (deep bass)
traffic	Generated	Filtered noise with amplitude modulation
cafe	Generated	Bandpass noise with random bursts
home	Generated	50/60Hz hum + random clicks
crowd	Generated	Bandpass noise with rhythmic modulation
mic	Generated	High-pass hiss + occasional pops

Configuration

Copy .env.example to .env and set your API keys:

cp .env.example .env

Full Pipeline Example

from py_speech_gen import (
    DatasetGenerator, Randomizer, NoiseMixer,
)

generator = DatasetGenerator(
    providers=[piper, elevenlabs],
    output_dir="./output",
    randomizer=Randomizer.preset("moderate", seed=42),
    noise_mixer=NoiseMixer(
        noise_types=["white", "traffic", "cafe"],
        snr_range=(15, 25),
        skip_prob=0.2,
        seed=42,
    ),
)

dataset = generator.generate_dataset(
    texts=["Sentence 1.", "Sentence 2."],
    dataset_name="demo",
    process_texts=True,
    texts_lang="en",
)

# Export full generation config for reproducibility
generator.export_config("./output/generation_config.json")

# Load config to reproduce
generator2 = DatasetGenerator.load_config(
    "./output/generation_config.json",
    providers=[piper, elevenlabs],
)

Custom Providers

You can create your own TTS provider by extending BaseProvider. See Custom Provider Tutorial.

Running Tests

pip install "py-speech-gen[dev]"
pytest tests/ -v