There are two ways to configure the application. Either by command-line flags or by a settings file. The settings file allows you to configure more options, but the command-line flags take precedence over the settings file.
On the first run, the application will create a settings.yaml file in the same folder as the executable with the default settings and the command-line flags that were used to start the application.
Edit the settings.yaml file with any text editor to change the settings.
Used settings file can be changed by using the --config command-line flag.
These take precedence to the settings file. But not all options are available as command-line flags.
| --flags | Default Value | Description |
|---|---|---|
--devices |
False | Print all available devices. |
--device_index |
-1 | Choose the input device to listen to and transcribe the audio from this device. '-1' = auto-select mic by default. |
--device_out_index |
-1 | the id of the output device (-1 = default active Speaker) |
--sample_rate |
16000 | Sample rate of the audio recording. |
--ai_device |
None | defines on which device the AI is loaded. can be cuda or cpu. auto-select by default |
--task |
transcribe | Choose between to transcribe or to translate the audio to English. |
--model |
small | Select model list. can be tiny, base, small, medium, large. where large models are not available for english only. |
--language |
None | language spoken in the audio, specify None to perform language detection |
--condition_on_previous_text |
False | Feed it the previous result to keep it consistent across recognition windows, but makes it more prone to getting stuck in a failure loop |
--energy |
300 | Energy level for mic to detect. |
--dynamic_energy |
False | Enable dynamic energy. |
--pause |
0.8 | Pause time before entry ends. |
--phrase_time_limit |
None | Phrase time limit (in seconds) before entry ends to break up long recognition tasks. |
--osc_ip |
127.0.0.1 | IP to send OSC messages to. Set to '0' to disable. (For VRChat this should mostly be 127.0.0.1) |
--osc_port |
9000 | Port to send OSC message to. ('9000' as default for VRChat) |
--osc_address |
/chatbox/input | The Address the OSC messages are send to. ('/chatbox/input' as default for VRChat) |
--osc_convert_ascii |
False | Convert Text to ASCII compatible when sending over OSC. |
--websocket_ip |
0 | IP where Websocket Server listens on. Set to '0' to disable. |
--websocket_port |
5000 | Port where Websocket Server listens on. |
--txt_translator |
NLLB200 | The Model the AI is loading for text translations. can be 'M2M100', 'NLLB200' or 'None'. |
--txt_translator_size |
small | The Model size of M2M100 or NLLB200 text translator is used. can be 'small', 'medium', 'large' for NLLB200, or 'small' or 'large' for M2M100. |
--txt_translator_device |
auto | The device used for M2M100 translation. can be 'auto', 'cuda' or 'cpu'. |
--ocr_window_name |
VRChat | Window name of the application for OCR translations. |
--open_browser |
False | Open default Browser with websocket-remote on start. (requires --websocket_ip to be set as well) |
--config |
None | Use the specified config file instead of the default 'settings.yaml' (relative to the current path) [overwrites without asking!!!] |
--verbose |
False | Whether to print verbose output. |
All possible options of the settings file.
Default name is settings.yaml, but can be customized with the --config option.
# audio settings
audio_api: "MME", # The name of the audio API. (MME, DirectSound, WASAPI)
audio_input_device: "", # audio input device name - used by whispering tiger UI to select audio input device by name
audio_output_device: "", # audio output device name - used by whispering tiger UI to select audio output device by name
device_index: None, # input device index for STT
# whisper settings
ai_device: null # can be null (auto), "cuda" or "cpu".
whisper_task: translate # Whisper A.I. Can do "transcribe" or "translate".
current_language: null # can be null (auto) or any Whisper supported language (improves accuracy if whisper does not have to detect the language).
model: small # Whisper model size. Can be "tiny", "base", "small", "medium" or "large".
condition_on_previous_text: true # if enabled, Whisper will condition on previous text (more prone to loops or getting stuck).
prompt_reset_on_temperature: 0.5, # after which temperature fallback step the prompt with the previous text should be reset (default value is 0.5)
energy: 300, # energy of audio volume to start whisper processing. Can be 0-?????
phrase_time_limit: 0, # time limit for Whisper to generate a phrase. (0 = no limit)
pause: 1.0, # pause between phrases.
initial_prompt: "" # initial prompt for Whisper to try to follow its style. for example "Umm, let me think like, hmm... Okay, here's what I'm, like, thinking." will give more filler words.
logprob_threshold: "-1.0", # log probability threshold for Whisper to treat as failed. (can be negative or positive).
no_speech_threshold: "0.6", # If the no_speech probability is higher than this value AND the average log probability over sampled tokens is below `logprob_threshold`, consider the segment as silent
length_penalty: 1.0,
beam_search_patience: 1.0,
repetition_penalty: 1.0, # penalize the score of previously generated tokens (set > 1 to penalize)
no_repeat_ngram_size: 0, # prevent repetitions of ngrams with this size
whisper_precision: "float32" # for original Whisper can be "float16" or "float32", for faster-whisper "default", "auto", "int8", "int8_float16", "int16", "float16", "float32".
stt_type: "faster_whisper", # can be "faster_whisper", "original_whisper", "speech_t5" or "seamless_m4t".
temperature_fallback: true # Set to False to disable temperature fallback which is the reason for some slowdowns, but decreases quality.
beam_size: 5 # Beam size for beam search. (higher = more accurate, but slower)
whisper_cpu_threads: 0 # Number of threads to use when running on CPU (4 by default)
whisper_num_workers: 1 # When transcribe() is called from multiple Python threads
vad_enabled: True, # Enable Voice activity detection (VAD)
vad_on_full_clip: False, # If enabled, an additional VAD check will be applied to the full clip, not just the frames.
vad_confidence_threshold: "0.4", # Voice activity detection (VAD) confidence threshold. Can be 0-1
vad_frames_per_buffer: 2000, # Voice activity detection (VAD) sample size (how many audio samples should be tested).
vad_thread_num: 1, # number of threads to use for VAD.
push_to_talk_key: "", # Push to talk key or key combination. (empty or None to disable)
word_timestamps: False, # if enabled, Whisper will add timestamps to the transcribed text.
faster_without_timestamps: False, # if enabled, faster whisper will only sample text tokens. (only when using stt_type=faster_whisper)
whisper_apply_voice_markers: False, # if enabled, Whisper will apply voice markers.
max_sentence_repetition: -1, # set max sentence repetition in result (-1 = disabled)
transcription_auto_save_file: "", # set to filepath to save transcriptions. (empty or None to disable)
transcription_auto_save_continous_text: False, # set to save continous text line instead of CSV
transcription_save_audio_dir: "", # set to filepath to save transcriptions wav files. (empty or None to disable)
silence_cutting_enabled: True,
silence_offset: -40.0,
max_silence_length: 30.0,
keep_silence_length: 0.20,
normalize_enabled: True,
normalize_lower_threshold: -24.0,
normalize_upper_threshold: -16.0,
normalize_gain_factor: 2.0,
denoise_audio: False, # if enabled, audio will be de-noised before processing.
denoise_audio_post_filter: False, # Enable post filter for some minor, extra noise reduction.
realtime: false # if enabled, Whisper will process audio in realtime.
realtime_whisper_model: '' # model used for realtime transcription. (empty for using same model as model setting)
realtime_whisper_precision: "float16" # precision used for realtime transcription model.
realtime_whisper_beam_size: 1 # beam size used for realtime transcription model.
realtime_temperature_fallback: false # Set to False to disable temperature fallback for realtime transcription. (see temperature_fallback setting)
realtime_frame_multiply: 15 # Only sends the audio clip to Whisper every X frames. (higher = less whisper updates and less processing time)
realtime_frequency_time: 1.0 # Only sends the audio clip to Whisper every X seconds. (higher = less whisper updates and less processing time)
# text translate settings
txt_translate: false # if enabled, pipes whisper A.I. results through text translator.
txt_translator_device: 'cuda' # can be "auto", "cuda" or "cpu".
src_lang: auto # source language for text translator.
trg_lang: fra_Latn # target language for text translator.
txt_romaji: false # if enabled, text translator will convert text to romaji.
txt_translator: NLLB200 # can be "NLLB200" or "M2M100".
txt_translator_size: small # for M2M100 model size: Can be "small" or "large", for NLLB200 model size: Can be "small", "medium", "large".
txt_translator_precision: float32 # for ctranwslate based text translators: can be "default", "auto", "int8", "int8_float16", "int16", "float16", "float32".
txt_translate_realtime: false # use text translator in realtime mode
# websocket settings
websocket_ip: 127.0.0.1
websocket_port: 5000
# OSC settings
osc_ip: '127.0.0.1'
osc_port: 9000
osc_address: /chatbox/input
osc_typing_indicator: true # Display typing indicator while processing audio
osc_convert_ascii: false
osc_chat_prefix: '' # Prefix for OSC messages. (is prepended in front of the OSC message)
osc_chat_limit: 144, # defines the maximum length of a chat message.
osc_time_limit: 15.0, # defines the time between OSC messages in seconds.
osc_scroll_time_limit: 1.5, # defines the scroll time limit for scrolling OSC messages. (only used when osc_send_type is set to "scroll")
osc_initial_time_limit: 15.0, # defines the initial time after the first message is send.
osc_scroll_size: 3, # defines the scroll size for scrolling OSC messages. (only used when osc_send_type is set to "scroll")
osc_max_scroll_size: 30, # defines the maximum scroll size for scrolling OSC messages. ~30 to scroll on only a single line (only used when osc_send_type is set to "scroll")
osc_send_type: "chunks", # defines the type of OSC messages are send. Can be "scroll", "full_or_scroll", "chunks" or "full". Where "scroll" sends the text scrollung until all is send, "full_or_scroll" to only scroll if it is too long, "chunks" sends the text in chunks and "full" sends the whole text at once.
osc_auto_processing_enabled: True, # Toggle auto sending of OSC messages on WhisperAI results. (not saved)
osc_type_transfer: "translation_result", # defines which type of data to send. Can be "source", "translation_result" or "both".
osc_type_transfer_split: " 🌐 ", # defines how source and translation results are split. (only used when osc_type_transfer is set to "both")
osc_delay_until_audio_playback: False, # if enabled, OSC messages will be delayed until audio playback starts. (if no TTS is used, this will prevent messages from being send.)
osc_delay_until_audio_playback_tag: "tts", # defines the tag used for detecting audio playback. (only used when osc_delay_until_audio_playback is enabled. Set empty to detect any audio playback)
osc_delay_timeout: 10, # defines the timeout for delayed OSC messages. (only used when osc_delay_until_audio_playback is enabled)
# OCR settings
ocr_lang: en # language for OCR image to text recognition.
ocr_window_name: VRChat # window name for OCR image to text recognition.
# TTS settings
tts_enabled: True, # enable TTS
tts_ai_device: "cuda", # can be "auto", "cuda" or "cpu".
tts_answer: True, # answer to whisper results
device_out_index: None, # output device index for TTS
tts_model: ["en", "v3_en"], # TTS language and model to use
tts_voice: "en_0", # TTS voice (one of silero tts voices, or "last" to use last used voice)
tts_prosody_rate: "" # TTS voice speed. Can be "x-slow", "slow", "medium", "fast", "x-fast" or "" for default.
tts_prosody_pitch: "" # TTS voice pitch. Can be "x-low", "low", "medium", "high", "x-high" or "" for default.
# plugin settings
plugins: {} # list of plugins to load.
plugin_settings: {} # settings for plugins.
plugin_timer_timeout: 15.0 # timeout for plugin timer in seconds. (Timer pause time after whisper event)
plugin_timer: 2.0 # time between plugin timer events in seconds.