src/config.py

"""Configuration, phrase banks and default catalogs for the dataset creator."""

from __future__ import annotations

from dataclasses import dataclass, field
from typing import List


# --------------------------------------------------------------------------- #
# Default phrase banks (target phrase: "Hey Android")
# --------------------------------------------------------------------------- #

DEFAULT_WAKE_PHRASES: List[str] = [
    "Hey Android",
    "Hey, Android",
    "Hello Android",
    "Okay Android",
    "OK Android",
    "Android",
]

DEFAULT_UNKNOWN_PHRASES: List[str] = [
    "Hey Andrew",
    "Hey Adrian",
    "Hey Andrea",
    "Hello phone",
    "Open settings",
    "Start recording",
    "Stop recording",
    "Good morning",
    "Where is my phone",
    "Androids are useful",
    "This is a test",
    "Turn on the light",
    "Turn off the light",
    "Increase volume",
    "Decrease volume",
    "Play music",
    "Pause music",
]

# The three primary keyword-spotting classes.
WAKE_LABEL = "hey_android"
UNKNOWN_LABEL = "unknown"
NOISE_LABEL = "background_noise"

NOISE_TYPES: List[str] = ["white", "pink", "brown", "hum", "fan", "cafe", "street"]


# --------------------------------------------------------------------------- #
# Piper voice catalog (free, downloaded from rhasspy/piper-voices on the Hub)
# --------------------------------------------------------------------------- #

@dataclass(frozen=True)
class PiperVoice:
    """A downloadable Piper voice model on the Hugging Face Hub."""

    voice_id: str
    repo_path: str  # path inside rhasspy/piper-voices (without extension)
    locale: str
    description: str


PIPER_VOICES: List[PiperVoice] = [
    PiperVoice(
        "en_US-lessac-medium",
        "en/en_US/lessac/medium/en_US-lessac-medium",
        "en-US",
        "English US",
    ),
    PiperVoice(
        "en_GB-alba-medium",
        "en/en_GB/alba/medium/en_GB-alba-medium",
        "en-GB",
        "English GB",
    ),
    PiperVoice(
        "en_GB-northern_english_male-medium",
        "en/en_GB/northern_english_male/medium/en_GB-northern_english_male-medium",
        "en-GB",
        "English GB Northern male",
    ),
    PiperVoice(
        "nl_NL-mls-medium",
        "nl/nl_NL/mls/medium/nl_NL-mls-medium",
        "nl-NL",
        "Dutch NL",
    ),
    PiperVoice(
        "de_DE-thorsten-medium",
        "de/de_DE/thorsten/medium/de_DE-thorsten-medium",
        "de-DE",
        "German",
    ),
    PiperVoice(
        "fr_FR-siwis-medium",
        "fr/fr_FR/siwis/medium/fr_FR-siwis-medium",
        "fr-FR",
        "French",
    ),
    PiperVoice(
        "es_ES-sharvard-medium",
        "es/es_ES/sharvard/medium/es_ES-sharvard-medium",
        "es-ES",
        "Spanish",
    ),
]


# --------------------------------------------------------------------------- #
# Dataset generation config
# --------------------------------------------------------------------------- #

@dataclass
class DatasetConfig:
    """All knobs controlling a single dataset build."""

    # Output
    out_dir: str = "output"
    dataset_name: str = "hey_android"

    # Audio
    sample_rate_hz: int = 16000
    duration_seconds: float = 2.0

    # Classes / phrases
    wake_label: str = WAKE_LABEL
    unknown_label: str = UNKNOWN_LABEL
    noise_label: str = NOISE_LABEL
    wake_phrases: List[str] = field(default_factory=lambda: list(DEFAULT_WAKE_PHRASES))
    unknown_phrases: List[str] = field(default_factory=lambda: list(DEFAULT_UNKNOWN_PHRASES))

    # Size controls
    base_repeats_per_phrase_per_voice: int = 1
    augmentations_per_speech_clip: int = 8
    background_noise_samples: int = 200
    max_piper_voices: int = 7
    max_gcp_voices_per_locale: int = 3

    # GCP voice selection
    gcp_language_prefixes: List[str] = field(
        default_factory=lambda: ["en", "nl", "de", "fr", "es"]
    )

    # Split
    test_ratio: float = 0.2

    # Reproducibility
    seed: int = 1337

    @property
    def target_samples(self) -> int:
        return int(self.sample_rate_hz * self.duration_seconds)