bark package¶
+Submodules¶
+bark.api module¶
+-
+
- +bark.api.generate_audio(text: str, history_prompt: str | None = None, text_temp: float = 0.7, waveform_temp: float = 0.7, silent: bool = False, output_full: bool = False)¶ +
Generate audio array from input text.
+-
+
- Parameters: +
-
+
text – text to be turned into audio
+history_prompt – history choice for audio cloning
+text_temp – generation temperature (1.0 more diverse, 0.0 more conservative)
+waveform_temp – generation temperature (1.0 more diverse, 0.0 more conservative)
+silent – disable progress bar
+output_full – return full generation to be used as a history prompt
+
+- Returns: +
numpy audio array at sample frequency 24khz
+
+
-
+
- +bark.api.save_as_prompt(filepath, full_generation)¶ +
-
+
- +bark.api.semantic_to_waveform(semantic_tokens: ndarray, history_prompt: str | None = None, temp: float = 0.7, silent: bool = False, output_full: bool = False)¶ +
Generate audio array from semantic input.
+-
+
- Parameters: +
-
+
semantic_tokens – semantic token output from text_to_semantic
+history_prompt – history choice for audio cloning
+temp – generation temperature (1.0 more diverse, 0.0 more conservative)
+silent – disable progress bar
+output_full – return full generation to be used as a history prompt
+
+- Returns: +
numpy audio array at sample frequency 24khz
+
+
-
+
- +bark.api.text_to_semantic(text: str, history_prompt: str | None = None, temp: float = 0.7, silent: bool = False)¶ +
Generate semantic array from text.
+-
+
- Parameters: +
-
+
text – text to be turned into audio
+history_prompt – history choice for audio cloning
+temp – generation temperature (1.0 more diverse, 0.0 more conservative)
+silent – disable progress bar
+
+- Returns: +
numpy semantic array to be fed into semantic_to_waveform
+
+
bark.generation module¶
+-
+
- +class bark.generation.InferenceContext(benchmark=False)¶ +
Bases:
+object
-
+
- +bark.generation.autocast()¶ +
-
+
- +bark.generation.clean_models(model_key=None)¶ +
-
+
- +bark.generation.codec_decode(fine_tokens)¶ +
Turn quantized audio codes into audio array using encodec.
+
-
+
- +bark.generation.generate_coarse(x_semantic, history_prompt=None, temp=0.7, top_k=None, top_p=None, silent=False, max_coarse_history=630, sliding_window_len=60, use_kv_caching=False)¶ +
Generate coarse audio codes from semantic tokens.
+
-
+
- +bark.generation.generate_fine(x_coarse_gen, history_prompt=None, temp=0.5, silent=True)¶ +
Generate full audio codes from coarse audio codes.
+
-
+
- +bark.generation.generate_text_semantic(text, history_prompt=None, temp=0.7, top_k=None, top_p=None, silent=False, min_eos_p=0.2, max_gen_duration_s=None, allow_early_stop=True, use_kv_caching=False)¶ +
Generate semantic tokens from text.
+
-
+
- +bark.generation.load_codec_model(use_gpu=True, force_reload=False)¶ +
-
+
- +bark.generation.load_model(use_gpu=True, use_small=False, force_reload=False, model_type='text', path=None)¶ +
-
+
- +bark.generation.preload_models(text_use_gpu=True, text_use_small=False, coarse_use_gpu=True, coarse_use_small=False, fine_use_gpu=True, fine_use_small=False, codec_use_gpu=True, force_reload=False, path=None)¶ +
Load all the necessary models for the pipeline.
+
bark.model module¶
+Much of this code is adapted from Andrej Karpathy’s NanoGPT +(https://github.com/karpathy/nanoGPT)
+-
+
- +class bark.model.Block(config, layer_idx)¶ +
Bases:
+Module-
+
- +forward(x, past_kv=None, use_cache=False)¶ +
Defines the computation performed at every call.
+Should be overridden by all subclasses.
+++Note
+Although the recipe for forward pass needs to be defined within +this function, one should call the
+Moduleinstance afterwards +instead of this since the former takes care of running the +registered hooks while the latter silently ignores them.
-
+
- +class bark.model.CausalSelfAttention(config)¶ +
Bases:
+Module-
+
- +forward(x, past_kv=None, use_cache=False)¶ +
Defines the computation performed at every call.
+Should be overridden by all subclasses.
+++Note
+Although the recipe for forward pass needs to be defined within +this function, one should call the
+Moduleinstance afterwards +instead of this since the former takes care of running the +registered hooks while the latter silently ignores them.
-
+
- +class bark.model.GPT(config)¶ +
Bases:
+Module-
+
- +forward(idx, merge_context=False, past_kv=None, position_ids=None, use_cache=False, labels=None)¶ +
Defines the computation performed at every call.
+Should be overridden by all subclasses.
+++Note
+Although the recipe for forward pass needs to be defined within +this function, one should call the
+Moduleinstance afterwards +instead of this since the former takes care of running the +registered hooks while the latter silently ignores them.
-
+
- +get_num_params(non_embedding=True)¶ +
Return the number of parameters in the model. +For non-embedding count (default), the position embeddings get subtracted. +The token embeddings would too, except due to the parameter sharing these +params are actually used as weights in the final layer, so we include them.
+
-
+
- +class bark.model.GPTConfig(block_size: int = 1024, input_vocab_size: int = 10048, output_vocab_size: int = 10048, n_layer: int = 12, n_head: int = 12, n_embd: int = 768, dropout: float = 0.0, bias: bool = True)¶ +
Bases:
+object-
+
- +bias: bool = True¶ +
-
+
- +block_size: int = 1024¶ +
-
+
- +dropout: float = 0.0¶ +
-
+
- +input_vocab_size: int = 10048¶ +
-
+
- +n_embd: int = 768¶ +
-
+
- +n_head: int = 12¶ +
-
+
- +n_layer: int = 12¶ +
-
+
- +output_vocab_size: int = 10048¶ +
-
+
- +class bark.model.LayerNorm(ndim, bias)¶ +
Bases:
+ModuleLayerNorm but with an optional bias. PyTorch doesn’t support simply bias=False
+-
+
- +forward(input)¶ +
Defines the computation performed at every call.
+Should be overridden by all subclasses.
+++Note
+Although the recipe for forward pass needs to be defined within +this function, one should call the
+Moduleinstance afterwards +instead of this since the former takes care of running the +registered hooks while the latter silently ignores them.
-
+
- +class bark.model.MLP(config)¶ +
Bases:
+Module-
+
- +forward(x)¶ +
Defines the computation performed at every call.
+Should be overridden by all subclasses.
+++Note
+Although the recipe for forward pass needs to be defined within +this function, one should call the
+Moduleinstance afterwards +instead of this since the former takes care of running the +registered hooks while the latter silently ignores them.
bark.model_fine module¶
+Much of this code is adapted from Andrej Karpathy’s NanoGPT +(https://github.com/karpathy/nanoGPT)
+-
+
- +class bark.model_fine.FineBlock(config)¶ +
Bases:
+Module-
+
- +forward(x)¶ +
Defines the computation performed at every call.
+Should be overridden by all subclasses.
+++Note
+Although the recipe for forward pass needs to be defined within +this function, one should call the
+Moduleinstance afterwards +instead of this since the former takes care of running the +registered hooks while the latter silently ignores them.
-
+
- +class bark.model_fine.FineGPT(config)¶ +
Bases:
+GPT-
+
- +forward(pred_idx, idx)¶ +
Defines the computation performed at every call.
+Should be overridden by all subclasses.
+++Note
+Although the recipe for forward pass needs to be defined within +this function, one should call the
+Moduleinstance afterwards +instead of this since the former takes care of running the +registered hooks while the latter silently ignores them.
-
+
- +get_num_params(non_embedding=True)¶ +
Return the number of parameters in the model. +For non-embedding count (default), the position embeddings get subtracted. +The token embeddings would too, except due to the parameter sharing these +params are actually used as weights in the final layer, so we include them.
+
-
+
- +class bark.model_fine.FineGPTConfig(block_size: int = 1024, input_vocab_size: int = 10048, output_vocab_size: int = 10048, n_layer: int = 12, n_head: int = 12, n_embd: int = 768, dropout: float = 0.0, bias: bool = True, n_codes_total: int = 8, n_codes_given: int = 1)¶ +
Bases:
+GPTConfig-
+
- +n_codes_given: int = 1¶ +
-
+
- +n_codes_total: int = 8¶ +
-
+
- +class bark.model_fine.NonCausalSelfAttention(config)¶ +
Bases:
+Module-
+
- +forward(x)¶ +
Defines the computation performed at every call.
+Should be overridden by all subclasses.
+++Note
+Although the recipe for forward pass needs to be defined within +this function, one should call the
+Moduleinstance afterwards +instead of this since the former takes care of running the +registered hooks while the latter silently ignores them.
