Whisper

class ctranslate2.models.Whisper

Implements the Whisper speech recognition model published by OpenAI.

See also

https://github.com/openai/whisper

Inherits from: pybind11_builtins.pybind11_object

Attributes:

Methods:

__init__(model_path: str, device: str = 'cpu', *, device_index: Union[int, List[int]] = 0, compute_type: Union[str, Dict[str, str]] = 'default', inter_threads: int = 1, intra_threads: int = 0, max_queued_batches: int = 0, flash_attention: bool = False, tensor_parallel: bool = False, files: object = None) → None

Initializes a Whisper model from a converted model.

Parameters

  • model_path – Path to the CTranslate2 model directory.

  • device – Device to use (possible values are: cpu, cuda, auto).

  • device_index – Device IDs where to place this model on.

  • compute_type – Model computation type or a dictionary mapping a device name to the computation type (possible values are: default, auto, int8, int8_float32, int8_float16, int8_bfloat16, int16, float16, bfloat16, float32).

  • inter_threads – Number of workers to allow executing multiple batches in parallel.

  • intra_threads – Number of OpenMP threads per worker (0 to use a default value).

  • max_queued_batches – Maximum numbers of batches in the worker queue (-1 for unlimited, 0 for an automatic value). When the queue is full, future requests will block until a free slot is available.

  • flash_attention – run model with flash attention 2 for self-attention layer

  • tensor_parallel – run model with tensor parallel mode

  • files – Load model files from the memory. This argument is a dictionary mapping file names to file contents as file-like or bytes objects. If this is set, model_path acts as an identifier for this model.

align(features: StorageView, start_sequence: List[int], text_tokens: List[List[int]], num_frames: Union[int, List[int]], *, median_filter_width: int = 7) → List[WhisperAlignmentResult]

Computes the alignments between the text tokens and the audio.

Parameters

  • features – Mel spectogram of the audio, as a float array with shape [batch_size, n_mels, chunk_length]. This method also accepts the encoded features returned by the method ctranslate2.models.Whisper.encode(), which have shape [batch_size, chunk_length // 2, d_model].

  • start_sequence – The start sequence tokens.

  • text_tokens – Batch of text tokens to align.

  • num_frames – Number of non padding frames in the features.

  • median_filter_width – Width of the median filter kernel.

Returns

A list of alignment results.

detect_language(features: StorageView) → List[List[Tuple[str, float]]]

Returns the probability of each language.

Parameters

features – Mel spectogram of the audio, as a float array with shape [batch_size, n_mels, chunk_length]. This method also accepts the encoded features returned by the method ctranslate2.models.Whisper.encode(), which have shape [batch_size, chunk_length // 2, d_model].

Returns

For each batch, a list of pairs (language, probability) ordered from best to worst probability.

Raises

RuntimeError – if the model is not multilingual.

encode(features: StorageView, to_cpu: bool = False) → StorageView

Encodes the input features.

Parameters

  • features – Mel spectogram of the audio, as a float array with shape [batch_size, n_mels, chunk_length].

  • to_cpu – Copy the encoder output to the CPU before returning the value.

Returns

The encoder output.

generate(features: StorageView, prompts: Union[List[List[str]], List[List[int]]], *, asynchronous: bool = False, beam_size: int = 5, patience: float = 1, num_hypotheses: int = 1, length_penalty: float = 1, repetition_penalty: float = 1, no_repeat_ngram_size: int = 0, max_length: int = 448, return_scores: bool = False, return_no_speech_prob: bool = False, max_initial_timestamp_index: int = 50, suppress_blank: bool = True, suppress_tokens: Optional[List[int]] = [- 1], sampling_topk: int = 1, sampling_temperature: float = 1) → Union[List[WhisperGenerationResult], List[WhisperGenerationResultAsync]]

Encodes the input features and generates from the given prompt.

Parameters

  • features – Mel spectogram of the audio, as a float array with shape [batch_size, n_mels, chunk_length]. This method also accepts the encoded features returned by the method ctranslate2.models.Whisper.encode(), which have shape [batch_size, chunk_length // 2, d_model].

  • prompts – Batch of initial string tokens or token IDs.

  • asynchronous – Run the model asynchronously.

  • beam_size – Beam size (1 for greedy search).

  • patience – Beam search patience factor, as described in https://arxiv.org/abs/2204.05424. The decoding will continue until beam_size*patience hypotheses are finished.

  • num_hypotheses – Number of hypotheses to return.

  • length_penalty – Exponential penalty applied to the length during beam search.

  • repetition_penalty – Penalty applied to the score of previously generated tokens (set > 1 to penalize).

  • no_repeat_ngram_size – Prevent repetitions of ngrams with this size (set 0 to disable).

  • max_length – Maximum generation length.

  • return_scores – Include the scores in the output.

  • return_no_speech_prob – Include the probability of the no speech token in the result.

  • max_initial_timestamp_index – Maximum index of the first predicted timestamp.

  • suppress_blank – Suppress blank outputs at the beginning of the sampling.

  • suppress_tokens – List of token IDs to suppress. -1 will suppress a default set of symbols as defined in the model config.json file.

  • sampling_topk – Randomly sample predictions from the top K candidates.

  • sampling_temperature – Sampling temperature to generate more random samples.

Returns

A list of generation results.

load_model(keep_cache: bool = False) → None

Loads the model back to the initial device.

Parameters

keep_cache – If True, the model cache in the CPU memory is not deleted if it exists.

unload_model(to_cpu: bool = False) → None

Unloads the model attached to this whisper but keep enough runtime context to quickly resume whisper on the initial device.

Parameters

to_cpu – If True, the model is moved to the CPU memory and not fully unloaded.

property compute_type

Computation type used by the model.

property device

Device this model is running on.

property device_index

List of device IDs where this model is running on.

property is_multilingual

Returns True if this model is multilingual.

property model_is_loaded

Whether the model is loaded on the initial device and ready to be used.

property n_mels

Returns dimension of mel input features.

property num_active_batches

Number of batches waiting to be processed or currently processed.

property num_languages

Returns the number of languages supported.

property num_queued_batches

Number of batches waiting to be processed.

property num_workers

Number of model workers backing this instance.

property tensor_parallel

Run model with tensor parallel mode.