Bark Supported Languages: English (en), German (de), Spanish (es), French (fr), Hindi (hi), Italian (it), Japanese (ja), Korean (ko), Polish (pl), Portuguese (pt), Russian (ru), Turkish (tr), Chinese, simplified (zh)

Text Temperature: This parameter affects how the model generates speech from text. A higher text temperature value makes the model's output more random, while a lower text temperature value makes the model's output more deterministic. In other words, with a high text temperature, the model is more likely to generate unusual or unexpected speech from a given text prompt. On the other hand, with a low text temperature, the model is more likely to stick closely to the most probable output.

Waveform Temperature: This parameter affects how the model generates the final audio waveform. A higher waveform temperature value introduces more randomness into the audio output, which might result in more unusual sounds or voice modulations. A lower waveform temperature, on the other hand, makes the audio output more predictable and consistent.

Adjusting audio speed and pitch: Changes to speed and pitch may cause a fair amount of echo and reverb in the output audio. Running the audio through a third-party AI audio tool may help remove echo or reverb. A library called librosa is used for manipulating the audio speed and pitch. The speed of the audio is adjusted using the librosa.effects.time_stretch function, which stretches or compresses the audio by a certain factor. If the speed parameter passed into the generate_voice function is not 1.0 (i.e., the speed of the audio needs to be changed), the audio is time-stretched by the given rate. For instance, if the speed is 2, the audio's duration will be halved, making it play twice as fast. The pitch of the audio is adjusted using the librosa.effects.pitch_shift function. This function shifts the pitch of the audio by a certain number of half-steps. If the pitch parameter passed into the generate_voice function is not 0 (i.e., the pitch of the audio needs to be changed), the pitch of the audio is shifted by the given number of half-steps. For instance, if the pitch is 2, the pitch of the audio will be increased by 2 half-steps.

Noise Reduction (NR): Reduce background noise (not as good as an AI enhanced cleaner and often difficult to tell impact to audio given the randomness of each Bark generated speech even with same settings, it also can't remove echoing or AI hallucination). Code Ref (bark_connector.py): If value of 'reduce_noise' is True, it triggers noise reduction on the generated audio using the noisereduce library. reduce_noise takes the audio data and the sample rate as parameters and returns the audio with reduced noise. If reduce_noise is False, no noise reduction is applied, and the original audio is used.

Remove Silence (RS): RRemove any extended pauses or silence (may not do much, was included for situations when generated voice contains long pauses for unknown reasons). Code Ref (bark_connector.py): If value of 'remove_silence' is True, it enables aggressive silence removal by setting the VAD (Voice Activity Detection) to level 3. The webrtcvad library is used for voice activity detection. If remove_silence is False, the VAD level is set to 0, which means no silence removal is applied. The sample rate also had to be reduced to 16000 from 24000 to get it to work with the webrtcvad library.