Config

Example Config File

example_config.json

{
    "do_classify": true,
    "filter_empty_transcript": true,
    "classifier": {
        "model": "bookbot/distil-wav2vec2-adult-child-cls-52m",
        "max_duration_s": 3.0
    },
    "transcriber": {
        "type": "wav2vec2",
        "model": "bookbot/wav2vec2-bookbot-en-lm",
        "return_timestamps": "word",
        "chunk_length_s": 30
    },
    "do_noise_classify": true,
    "noise_classifier": {
        "model": "bookbot/distil-ast-audioset",
        "minimum_empty_duration": 0.3,
        "threshold": 0.2
    },
    "segmenter": {
        "type": "word_overlap",
        "minimum_chunk_duration": 1.0
    }
}

---

speechline.config.NoiseClassifierConfig dataclass

Noise classifier config.

Parameters:

Name	Type	Description	Default
model	str	HuggingFace Hub model hub checkpoint.	required
min_empty_duration	float	Minimum non-transcribed segment duration to be segmented, and passed to noise classifier. Defaults to 1.0 seconds.	required
threshold	float	The probability threshold for the multi label classification. Defaults to 0.3.	0.3
batch_size	int	Batch size during inference. Defaults to 1.	1

Source code in speechline/config.py

@dataclass
class NoiseClassifierConfig:
    """
    Noise classifier config.

    Args:
        model (str):
            HuggingFace Hub model hub checkpoint.
        min_empty_duration (float, optional):
            Minimum non-transcribed segment duration to be segmented,
            and passed to noise classifier.
            Defaults to `1.0` seconds.
        threshold (float, optional):
            The probability threshold for the multi label classification.
            Defaults to `0.3`.
        batch_size (int, optional):
            Batch size during inference. Defaults to `1`.

    """

    model: str
    minimum_empty_duration: float = 1.0
    threshold: float = 0.3
    batch_size: int = 1

speechline.config.ClassifierConfig dataclass

Audio classifier config.

Parameters:

Name	Type	Description	Default
model	str	HuggingFace Hub model hub checkpoint.	required
max_duration_s	float	Maximum audio duration for padding. Defaults to 3.0 seconds.	3.0
batch_size	int	Batch size during inference. Defaults to 1.	1

Source code in speechline/config.py

@dataclass
class ClassifierConfig:
    """
    Audio classifier config.

    Args:
        model (str):
            HuggingFace Hub model hub checkpoint.
        max_duration_s (float, optional):
            Maximum audio duration for padding. Defaults to `3.0` seconds.
        batch_size (int, optional):
            Batch size during inference. Defaults to `1`.
    """

    model: str
    max_duration_s: float = 3.0
    batch_size: int = 1

speechline.config.TranscriberConfig dataclass

Audio transcriber config.

Parameters:

Name	Type	Description	Default
type	str	Transcriber model architecture type.	required
model	str	HuggingFace Hub model hub checkpoint (not required for 'gentle').	None
return_timestamps	Union[str, bool]	return_timestamps argument in AutomaticSpeechRecognitionPipeline's __call__ method. Use "char" for CTC-based models and True for Whisper-based models.	None
chunk_length_s	int	Audio chunk length in seconds.	None
torch_dtype	str	Torch dtype for model weights (e.g., 'float16'). Used by Canary transcriber.	None
validate_alignment	bool	Enable alignment validation mode. Defaults to False.	False
token_confidence_threshold	float	Minimum confidence for token acceptance (0-1). Defaults to 0.7.	0.7
min_alignment_ratio	float	Minimum fraction of tokens that must align (0-1). Defaults to 0.8.	0.8
nfa_model	str	NeMo model for forced alignment. Defaults to "nvidia/parakeet-ctc-1.1b".	'nvidia/parakeet-ctc-1.1b'
gentle_path	str	Path to Gentle installation. Only used when type is "gentle". Defaults to "/mnt/Projects/Projects/AudioProcessing/gentle".	'/mnt/4090_projects/Projects/AudioProcessing/gentle'
output_phonemes	bool	Include phoneme sequences in Gentle output. Defaults to True.	True
output_word_boundaries	bool	Include word boundary timestamps in Gentle output. Defaults to True.	True

Source code in speechline/config.py

@dataclass
class TranscriberConfig:
    """
    Audio transcriber config.

    Args:
        type (str):
            Transcriber model architecture type.
        model (str):
            HuggingFace Hub model hub checkpoint (not required for 'gentle').
        return_timestamps (Union[str, bool]):
            `return_timestamps` argument in `AutomaticSpeechRecognitionPipeline`'s
            `__call__` method. Use `"char"` for CTC-based models and
            `True` for Whisper-based models.
        chunk_length_s (int):
            Audio chunk length in seconds.
        torch_dtype (str, optional):
            Torch dtype for model weights (e.g., 'float16'). Used by Canary transcriber.
        validate_alignment (bool, optional):
            Enable alignment validation mode. Defaults to False.
        token_confidence_threshold (float, optional):
            Minimum confidence for token acceptance (0-1). Defaults to 0.7.
        min_alignment_ratio (float, optional):
            Minimum fraction of tokens that must align (0-1). Defaults to 0.8.
        nfa_model (str, optional):
            NeMo model for forced alignment. Defaults to "nvidia/parakeet-ctc-1.1b".
        gentle_path (str, optional):
            Path to Gentle installation. Only used when type is "gentle".
            Defaults to "/mnt/Projects/Projects/AudioProcessing/gentle".
        output_phonemes (bool, optional):
            Include phoneme sequences in Gentle output. Defaults to True.
        output_word_boundaries (bool, optional):
            Include word boundary timestamps in Gentle output. Defaults to True.
    """

    type: str
    model: str = None
    return_timestamps: Union[str, bool] = None
    chunk_length_s: Optional[int] = None
    transcriber_device: str = "cuda"
    torch_dtype: Optional[str] = None
    validate_alignment: bool = False
    token_confidence_threshold: float = 0.7
    min_alignment_ratio: float = 0.8
    nfa_model: str = "nvidia/parakeet-ctc-1.1b"
    gentle_path: str = "/mnt/4090_projects/Projects/AudioProcessing/gentle"
    output_phonemes: bool = True
    output_word_boundaries: bool = True

    def __post_init__(self):
        SUPPORTED_MODELS = {"wav2vec2", "whisper", "parakeet", "parakeet_tdt", "canary", "gentle"}
        WAV2VEC_TIMESTAMPS = {"word", "char"}
        PARAKEET_TIMESTAMPS = {"word"}
        PARAKEET_TDT_TIMESTAMPS = {"word", "char"}
        GENTLE_TIMESTAMPS = {"word"}

        if self.type not in SUPPORTED_MODELS:
            raise ValueError(f"Transcriber of type {self.type} is not yet supported!")

        # Gentle has different requirements
        if self.type == "gentle":
            # Gentle doesn't need model checkpoint
            if self.return_timestamps is None:
                self.return_timestamps = "word"
            if self.return_timestamps not in GENTLE_TIMESTAMPS:
                raise ValueError("gentle only supports `'word'` timestamps!")
            return

        # All other transcribers require model
        if self.model is None:
            raise ValueError(f"model is required for {self.type} transcriber")

        if self.type == "wav2vec2" and self.return_timestamps not in WAV2VEC_TIMESTAMPS:
            raise ValueError("wav2vec2 only supports `'word'` or `'char'` timestamps!")
        elif self.type == "parakeet" and self.return_timestamps not in PARAKEET_TIMESTAMPS:
            raise ValueError("parakeet only supports `word` timestamps!")
        elif self.type == "parakeet_tdt" and self.return_timestamps not in PARAKEET_TDT_TIMESTAMPS:
            raise ValueError("parakeet_tdt only supports `'word'` or `'char'` timestamps!")
        elif self.type in {"whisper", "canary"} and self.return_timestamps is not True:
            raise ValueError(f"{self.type} only supports `True` timestamps!")

        # Add validation for chunk_length_s requirement
        if self.type in {"wav2vec2", "whisper", "canary", "parakeet_tdt"} and self.chunk_length_s is None:
            raise ValueError(f"chunk_length_s is required for {self.type} models")

speechline.config.SegmenterConfig dataclass

Audio segmenter config.

Parameters:

Name	Type	Description	Default
silence_duration	float	Minimum in-between silence duration (in seconds) to consider as gaps. Defaults to 3.0 seconds.	0.0
minimum_chunk_duration	float	Minimum chunk duration (in seconds) to be exported. Defaults to 0.2 second.	0.2
lexicon_path	str	Path to lexicon file. Defaults to None.	None
keep_whitespace	bool	Whether to keep whitespace in transcript. Defaults to False.	False

Source code in speechline/config.py

@dataclass
class SegmenterConfig:
    """
    Audio segmenter config.

    Args:
        silence_duration (float, optional):
            Minimum in-between silence duration (in seconds) to consider as gaps.
            Defaults to `3.0` seconds.
        minimum_chunk_duration (float, optional):
            Minimum chunk duration (in seconds) to be exported.
            Defaults to 0.2 second.
        lexicon_path (str, optional):
            Path to lexicon file. Defaults to `None`.
        keep_whitespace (bool, optional):
            Whether to keep whitespace in transcript. Defaults to `False`.
    """

    type: str
    silence_duration: float = 0.0
    minimum_chunk_duration: float = 0.2
    lexicon_path: str = None
    keep_whitespace: bool = False
    segment_with_ground_truth: bool = False

    def __post_init__(self):
        SUPPORTED_TYPES = {"silence", "word_overlap", "phoneme_overlap"}

        if self.type not in SUPPORTED_TYPES:
            raise ValueError(f"Segmenter of type {self.type} is not yet supported!")

speechline.config.Config dataclass

Main SpeechLine config, contains all other subconfigs.

Parameters:

Name	Type	Description	Default
path	str	Path to JSON config file.	required

Source code in speechline/config.py

@dataclass
class Config:
    """
    Main SpeechLine config, contains all other subconfigs.

    Args:
        path (str):
            Path to JSON config file.
    """

    path: str

    def __post_init__(self):
        config = json.load(open(self.path))
        self.do_classify = config.get("do_classify", False)
        self.do_noise_classify = config.get("do_noise_classify", False)
        self.filter_empty_transcript = config.get("filter_empty_transcript", False)
        self.audio_extension = config.get("audio_extension", "wav")
        self.folder_filter = config.get("folder_filter", None)

        if self.do_classify:
            self.classifier = ClassifierConfig(**config["classifier"])

        if self.do_noise_classify:
            self.noise_classifier = NoiseClassifierConfig(**config["noise_classifier"])

        self.transcriber = TranscriberConfig(**config["transcriber"])
        self.segmenter = SegmenterConfig(**config["segmenter"])