Merge pull request #200 from jhj0517/feature/add-parameter

app.py

@@ -1,5 +1,6 @@
 import os
 import argparse
+import gradio as gr
 
 from modules.whisper.whisper_Inference import WhisperInference
 from modules.whisper.faster_whisper_inference import FasterWhisperInference
@@ -59,20 +60,126 @@ class App:
             )
         return whisper_inf
 
-    @staticmethod
-    def open_folder(folder_path: str):
-        if os.path.exists(folder_path):
-            os.system(f"start {folder_path}")
-        else:
-            print(f"The folder {folder_path} does not exist.")
+    def create_whisper_parameters(self):
+        with gr.Row():
+            dd_model = gr.Dropdown(choices=self.whisper_inf.available_models, value="large-v2",
+                                   label="Model")
+            dd_lang = gr.Dropdown(choices=["Automatic Detection"] + self.whisper_inf.available_langs,
+                                  value="Automatic Detection", label="Language")
+            dd_file_format = gr.Dropdown(choices=["SRT", "WebVTT", "txt"], value="SRT", label="File Format")
+        with gr.Row():
+            cb_translate = gr.Checkbox(value=False, label="Translate to English?", interactive=True)
+        with gr.Row():
+            cb_timestamp = gr.Checkbox(value=True, label="Add a timestamp to the end of the filename",
+                                       interactive=True)
+        with gr.Accordion("Advanced Parameters", open=False):
+            nb_beam_size = gr.Number(label="Beam Size", value=1, precision=0, interactive=True,
+                                     info="Beam size to use for decoding.")
+            nb_log_prob_threshold = gr.Number(label="Log Probability Threshold", value=-1.0, interactive=True,
+                                              info="If the average log probability over sampled tokens is below this value, treat as failed.")
+            nb_no_speech_threshold = gr.Number(label="No Speech Threshold", value=0.6, interactive=True,
+                                               info="If the no speech probability is higher than this value AND the average log probability over sampled tokens is below 'Log Prob Threshold', consider the segment as silent.")
+            dd_compute_type = gr.Dropdown(label="Compute Type", choices=self.whisper_inf.available_compute_types,
+                                          value=self.whisper_inf.current_compute_type, interactive=True,
+                                          info="Select the type of computation to perform.")
+            nb_best_of = gr.Number(label="Best Of", value=5, interactive=True,
+                                   info="Number of candidates when sampling with non-zero temperature.")
+            nb_patience = gr.Number(label="Patience", value=1, interactive=True,
+                                    info="Beam search patience factor.")
+            cb_condition_on_previous_text = gr.Checkbox(label="Condition On Previous Text", value=True,
+                                                        interactive=True,
+                                                        info="Condition on previous text during decoding.")
+            sld_prompt_reset_on_temperature = gr.Slider(label="Prompt Reset On Temperature", value=0.5,
+                                                        minimum=0, maximum=1, step=0.01, interactive=True,
+                                                        info="Resets prompt if temperature is above this value."
+                                                             " Arg has effect only if 'Condition On Previous Text' is True.")
+            tb_initial_prompt = gr.Textbox(label="Initial Prompt", value=None, interactive=True,
+                                           info="Initial prompt to use for decoding.")
+            sd_temperature = gr.Slider(label="Temperature", value=0, step=0.01, maximum=1.0, interactive=True,
+                                       info="Temperature for sampling. It can be a tuple of temperatures, which will be successively used upon failures according to either `Compression Ratio Threshold` or `Log Prob Threshold`.")
+            nb_compression_ratio_threshold = gr.Number(label="Compression Ratio Threshold", value=2.4, interactive=True,
+                                                       info="If the gzip compression ratio is above this value, treat as failed.")
+            with gr.Group(visible=isinstance(self.whisper_inf, FasterWhisperInference)):
+                nb_length_penalty = gr.Number(label="Length Penalty", value=1,
+                                              info="Exponential length penalty constant.")
+                nb_repetition_penalty = gr.Number(label="Repetition Penalty", value=1,
+                                                  info="Penalty applied to the score of previously generated tokens (set > 1 to penalize).")
+                nb_no_repeat_ngram_size = gr.Number(label="No Repeat N-gram Size", value=0, precision=0,
+                                                    info="Prevent repetitions of n-grams with this size (set 0 to disable).")
+                tb_prefix = gr.Textbox(label="Prefix", value=lambda: None,
+                                       info="Optional text to provide as a prefix for the first window.")
+                cb_suppress_blank = gr.Checkbox(label="Suppress Blank", value=True,
+                                                info="Suppress blank outputs at the beginning of the sampling.")
+                tb_suppress_tokens = gr.Textbox(label="Suppress Tokens", value="[-1]",
+                                                info="List of token IDs to suppress. -1 will suppress a default set of symbols as defined in the model config.json file.")
+                nb_max_initial_timestamp = gr.Number(label="Max Initial Timestamp", value=1.0,
+                                                     info="The initial timestamp cannot be later than this.")
+                cb_word_timestamps = gr.Checkbox(label="Word Timestamps", value=False,
+                                                 info="Extract word-level timestamps using the cross-attention pattern and dynamic time warping, and include the timestamps for each word in each segment.")
+                tb_prepend_punctuations = gr.Textbox(label="Prepend Punctuations", value="\"'“¿([{-",
+                                                     info="If 'Word Timestamps' is True, merge these punctuation symbols with the next word.")
+                tb_append_punctuations = gr.Textbox(label="Append Punctuations", value="\"'.。,，!！?？:：”)]}、",
+                                                    info="If 'Word Timestamps' is True, merge these punctuation symbols with the previous word.")
+                nb_max_new_tokens = gr.Number(label="Max New Tokens", value=lambda: None, precision=0,
+                                              info="Maximum number of new tokens to generate per-chunk. If not set, the maximum will be set by the default max_length.")
+                nb_chunk_length = gr.Number(label="Chunk Length", value=lambda: None, precision=0,
+                                            info="The length of audio segments. If it is not None, it will overwrite the default chunk_length of the FeatureExtractor.")
+                nb_hallucination_silence_threshold = gr.Number(label="Hallucination Silence Threshold (sec)",
+                                                               value=lambda: None,
+                                                               info="When 'Word Timestamps' is True, skip silent periods longer than this threshold (in seconds) when a possible hallucination is detected.")
+                tb_hotwords = gr.Textbox(label="Hotwords", value=None,
+                                         info="Hotwords/hint phrases to provide the model with. Has no effect if prefix is not None.")
+                nb_language_detection_threshold = gr.Number(label="Language Detection Threshold", value=None,
+                                                            info="If the maximum probability of the language tokens is higher than this value, the language is detected.")
+                nb_language_detection_segments = gr.Number(label="Language Detection Segments", value=1, precision=0,
+                                                           info="Number of segments to consider for the language detection.")
+            with gr.Group(visible=isinstance(self.whisper_inf, InsanelyFastWhisperInference)):
+                nb_chunk_length_s = gr.Number(label="Chunk Lengths (sec)", value=30, precision=0)
+                nb_batch_size = gr.Number(label="Batch Size", value=24, precision=0)
 
-    @staticmethod
-    def on_change_models(model_size: str):
-        translatable_model = ["large", "large-v1", "large-v2", "large-v3"]
-        if model_size not in translatable_model:
-            return gr.Checkbox(visible=False, value=False, interactive=False)
-        else:
-            return gr.Checkbox(visible=True, value=False, label="Translate to English?", interactive=True)
+        with gr.Accordion("VAD", open=False):
+            cb_vad_filter = gr.Checkbox(label="Enable Silero VAD Filter", value=False, interactive=True)
+            sd_threshold = gr.Slider(minimum=0.0, maximum=1.0, step=0.01, label="Speech Threshold", value=0.5,
+                                     info="Lower it to be more sensitive to small sounds.")
+            nb_min_speech_duration_ms = gr.Number(label="Minimum Speech Duration (ms)", precision=0, value=250)
+            nb_max_speech_duration_s = gr.Number(label="Maximum Speech Duration (s)", value=9999)
+            nb_min_silence_duration_ms = gr.Number(label="Minimum Silence Duration (ms)", precision=0, value=2000)
+            nb_speech_pad_ms = gr.Number(label="Speech Padding (ms)", precision=0, value=400)
+
+        with gr.Accordion("Diarization", open=False):
+            cb_diarize = gr.Checkbox(label="Enable Diarization")
+            tb_hf_token = gr.Text(label="HuggingFace Token", value="",
+                                  info="This is only needed the first time you download the model. If you already have models, you don't need to enter. To download the model, you must manually go to \"https://huggingface.co/pyannote/speaker-diarization-3.1\" and agree to their requirement.")
+            dd_diarization_device = gr.Dropdown(label="Device",
+                                                choices=self.whisper_inf.diarizer.get_available_device(),
+                                                value=self.whisper_inf.diarizer.get_device())
+
+        dd_model.change(fn=self.on_change_models, inputs=[dd_model], outputs=[cb_translate])
+
+        return (
+            WhisperParameters(
+                model_size=dd_model, lang=dd_lang, is_translate=cb_translate, beam_size=nb_beam_size,
+                log_prob_threshold=nb_log_prob_threshold, no_speech_threshold=nb_no_speech_threshold,
+                compute_type=dd_compute_type, best_of=nb_best_of, patience=nb_patience,
+                condition_on_previous_text=cb_condition_on_previous_text, initial_prompt=tb_initial_prompt,
+                temperature=sd_temperature, compression_ratio_threshold=nb_compression_ratio_threshold,
+                vad_filter=cb_vad_filter, threshold=sd_threshold, min_speech_duration_ms=nb_min_speech_duration_ms,
+                max_speech_duration_s=nb_max_speech_duration_s, min_silence_duration_ms=nb_min_silence_duration_ms,
+                speech_pad_ms=nb_speech_pad_ms, chunk_length_s=nb_chunk_length_s, batch_size=nb_batch_size,
+                is_diarize=cb_diarize, hf_token=tb_hf_token, diarization_device=dd_diarization_device,
+                length_penalty=nb_length_penalty, repetition_penalty=nb_repetition_penalty,
+                no_repeat_ngram_size=nb_no_repeat_ngram_size, prefix=tb_prefix, suppress_blank=cb_suppress_blank,
+                suppress_tokens=tb_suppress_tokens, max_initial_timestamp=nb_max_initial_timestamp,
+                word_timestamps=cb_word_timestamps, prepend_punctuations=tb_prepend_punctuations,
+                append_punctuations=tb_append_punctuations, max_new_tokens=nb_max_new_tokens, chunk_length=nb_chunk_length,
+                hallucination_silence_threshold=nb_hallucination_silence_threshold, hotwords=tb_hotwords,
+                language_detection_threshold=nb_language_detection_threshold,
+                language_detection_segments=nb_language_detection_segments,
+                prompt_reset_on_temperature=sld_prompt_reset_on_temperature
+            ),
+            dd_file_format,
+            cb_timestamp
+        )
 
     def launch(self):
         with self.app:
@@ -84,47 +191,13 @@ class App:
                     with gr.Column():
                         input_file = gr.Files(type="filepath", label="Upload File here")
                         tb_input_folder = gr.Textbox(label="Input Folder Path (Optional)",
-                                                     info="Optional: Specify the folder path where the input files are located, if you prefer to use local files instead of uploading them." 
+                                                     info="Optional: Specify the folder path where the input files are located, if you prefer to use local files instead of uploading them."
                                                           " Leave this field empty if you do not wish to use a local path.",
                                                      visible=self.args.colab,
                                                      value="")
-                    with gr.Row():
-                        dd_model = gr.Dropdown(choices=self.whisper_inf.available_models, value="large-v2",
-                                               label="Model")
-                        dd_lang = gr.Dropdown(choices=["Automatic Detection"] + self.whisper_inf.available_langs,
-                                              value="Automatic Detection", label="Language")
-                        dd_file_format = gr.Dropdown(["SRT", "WebVTT", "txt"], value="SRT", label="File Format")
-                    with gr.Row():
-                        cb_translate = gr.Checkbox(value=False, label="Translate to English?", interactive=True)
-                    with gr.Row():
-                        cb_timestamp = gr.Checkbox(value=True, label="Add a timestamp to the end of the filename", interactive=True)
-                    with gr.Accordion("Advanced Parameters", open=False):
-                        nb_beam_size = gr.Number(label="Beam Size", value=1, precision=0, interactive=True)
-                        nb_log_prob_threshold = gr.Number(label="Log Probability Threshold", value=-1.0, interactive=True)
-                        nb_no_speech_threshold = gr.Number(label="No Speech Threshold", value=0.6, interactive=True)
-                        dd_compute_type = gr.Dropdown(label="Compute Type", choices=self.whisper_inf.available_compute_types, value=self.whisper_inf.current_compute_type, interactive=True)
-                        nb_best_of = gr.Number(label="Best Of", value=5, interactive=True)
-                        nb_patience = gr.Number(label="Patience", value=1, interactive=True)
-                        cb_condition_on_previous_text = gr.Checkbox(label="Condition On Previous Text", value=True, interactive=True)
-                        tb_initial_prompt = gr.Textbox(label="Initial Prompt", value=None, interactive=True)
-                        sd_temperature = gr.Slider(label="Temperature", value=0, step=0.01, maximum=1.0, interactive=True)
-                        nb_compression_ratio_threshold = gr.Number(label="Compression Ratio Threshold", value=2.4, interactive=True)
-                    with gr.Accordion("VAD", open=False):
-                        cb_vad_filter = gr.Checkbox(label="Enable Silero VAD Filter", value=False, interactive=True)
-                        sd_threshold = gr.Slider(minimum=0.0, maximum=1.0, step=0.01, label="Speech Threshold", value=0.5, info="Lower it to be more sensitive to small sounds.")
-                        nb_min_speech_duration_ms = gr.Number(label="Minimum Speech Duration (ms)", precision=0, value=250)
-                        nb_max_speech_duration_s = gr.Number(label="Maximum Speech Duration (s)", value=9999)
-                        nb_min_silence_duration_ms = gr.Number(label="Minimum Silence Duration (ms)", precision=0, value=2000)
-                        nb_speech_pad_ms = gr.Number(label="Speech Padding (ms)", precision=0, value=400)
-                    with gr.Accordion("Diarization", open=False):
-                        cb_diarize = gr.Checkbox(label="Enable Diarization")
-                        tb_hf_token = gr.Text(label="HuggingFace Token", value="",
-                                              info="This is only needed the first time you download the model. If you already have models, you don't need to enter. "
-                                                    "To download the model, you must manually go to \"https://huggingface.co/pyannote/speaker-diarization-3.1\" and agree to their requirement.")
-                        dd_diarization_device = gr.Dropdown(label="Device", choices=self.whisper_inf.diarizer.get_available_device(), value=self.whisper_inf.diarizer.get_device())
-                    with gr.Accordion("Insanely Fast Whisper Parameters", open=False, visible=isinstance(self.whisper_inf, InsanelyFastWhisperInference)):
-                        nb_chunk_length_s = gr.Number(label="Chunk Lengths (sec)", value=30, precision=0)
-                        nb_batch_size = gr.Number(label="Batch Size", value=24, precision=0)
+
+                    whisper_params, dd_file_format, cb_timestamp = self.create_whisper_parameters()
+
                     with gr.Row():
                         btn_run = gr.Button("GENERATE SUBTITLE FILE", variant="primary")
                     with gr.Row():
@@ -133,36 +206,10 @@ class App:
                         btn_openfolder = gr.Button('📂', scale=1)
 
                     params = [input_file, tb_input_folder, dd_file_format, cb_timestamp]
-                    whisper_params = WhisperParameters(model_size=dd_model,
-                                                       lang=dd_lang,
-                                                       is_translate=cb_translate,
-                                                       beam_size=nb_beam_size,
-                                                       log_prob_threshold=nb_log_prob_threshold,
-                                                       no_speech_threshold=nb_no_speech_threshold,
-                                                       compute_type=dd_compute_type,
-                                                       best_of=nb_best_of,
-                                                       patience=nb_patience,
-                                                       condition_on_previous_text=cb_condition_on_previous_text,
-                                                       initial_prompt=tb_initial_prompt,
-                                                       temperature=sd_temperature,
-                                                       compression_ratio_threshold=nb_compression_ratio_threshold,
-                                                       vad_filter=cb_vad_filter,
-                                                       threshold=sd_threshold,
-                                                       min_speech_duration_ms=nb_min_speech_duration_ms,
-                                                       max_speech_duration_s=nb_max_speech_duration_s,
-                                                       min_silence_duration_ms=nb_min_silence_duration_ms,
-                                                       speech_pad_ms=nb_speech_pad_ms,
-                                                       chunk_length_s=nb_chunk_length_s,
-                                                       batch_size=nb_batch_size,
-                                                       is_diarize=cb_diarize,
-                                                       hf_token=tb_hf_token,
-                                                       diarization_device=dd_diarization_device)
-
                     btn_run.click(fn=self.whisper_inf.transcribe_file,
                                   inputs=params + whisper_params.as_list(),
                                   outputs=[tb_indicator, files_subtitles])
                     btn_openfolder.click(fn=lambda: self.open_folder("outputs"), inputs=None, outputs=None)
-                    dd_model.change(fn=self.on_change_models, inputs=[dd_model], outputs=[cb_translate])
 
                 with gr.TabItem("Youtube"):  # tab2
                     with gr.Row():
@@ -173,45 +220,9 @@ class App:
                         with gr.Column():
                             tb_title = gr.Label(label="Youtube Title")
                             tb_description = gr.Textbox(label="Youtube Description", max_lines=15)
-                    with gr.Row():
-                        dd_model = gr.Dropdown(choices=self.whisper_inf.available_models, value="large-v2",
-                                               label="Model")
-                        dd_lang = gr.Dropdown(choices=["Automatic Detection"] + self.whisper_inf.available_langs,
-                                              value="Automatic Detection", label="Language")
-                        dd_file_format = gr.Dropdown(choices=["SRT", "WebVTT", "txt"], value="SRT", label="File Format")
-                    with gr.Row():
-                        cb_translate = gr.Checkbox(value=False, label="Translate to English?", interactive=True)
-                    with gr.Row():
-                        cb_timestamp = gr.Checkbox(value=True, label="Add a timestamp to the end of the filename",
-                                                   interactive=True)
-                    with gr.Accordion("Advanced Parameters", open=False):
-                        nb_beam_size = gr.Number(label="Beam Size", value=1, precision=0, interactive=True)
-                        nb_log_prob_threshold = gr.Number(label="Log Probability Threshold", value=-1.0, interactive=True)
-                        nb_no_speech_threshold = gr.Number(label="No Speech Threshold", value=0.6, interactive=True)
-                        dd_compute_type = gr.Dropdown(label="Compute Type", choices=self.whisper_inf.available_compute_types, value=self.whisper_inf.current_compute_type, interactive=True)
-                        nb_best_of = gr.Number(label="Best Of", value=5, interactive=True)
-                        nb_patience = gr.Number(label="Patience", value=1, interactive=True)
-                        cb_condition_on_previous_text = gr.Checkbox(label="Condition On Previous Text", value=True, interactive=True)
-                        tb_initial_prompt = gr.Textbox(label="Initial Prompt", value=None, interactive=True)
-                        sd_temperature = gr.Slider(label="Temperature", value=0, step=0.01, maximum=1.0, interactive=True)
-                        nb_compression_ratio_threshold = gr.Number(label="Compression Ratio Threshold", value=2.4, interactive=True)
-                    with gr.Accordion("VAD", open=False):
-                        cb_vad_filter = gr.Checkbox(label="Enable Silero VAD Filter", value=False, interactive=True)
-                        sd_threshold = gr.Slider(minimum=0.0, maximum=1.0, step=0.01, label="Speech Threshold", value=0.5, info="Lower it to be more sensitive to small sounds.")
-                        nb_min_speech_duration_ms = gr.Number(label="Minimum Speech Duration (ms)", precision=0, value=250)
-                        nb_max_speech_duration_s = gr.Number(label="Maximum Speech Duration (s)", value=9999)
-                        nb_min_silence_duration_ms = gr.Number(label="Minimum Silence Duration (ms)", precision=0, value=2000)
-                        nb_speech_pad_ms = gr.Number(label="Speech Padding (ms)", precision=0, value=400)
-                    with gr.Accordion("Diarization", open=False):
-                        cb_diarize = gr.Checkbox(label="Enable Diarization")
-                        tb_hf_token = gr.Text(label="HuggingFace Token", value="",
-                                              info="This is only needed the first time you download the model. If you already have models, you don't need to enter. "
-                                                    "To download the model, you must manually go to \"https://huggingface.co/pyannote/speaker-diarization-3.1\" and agree to their requirement.")
-                        dd_diarization_device = gr.Dropdown(label="Device", choices=self.whisper_inf.diarizer.get_available_device(), value=self.whisper_inf.diarizer.get_device())
-                    with gr.Accordion("Insanely Fast Whisper Parameters", open=False,
-                                      visible=isinstance(self.whisper_inf, InsanelyFastWhisperInference)):
-                        nb_chunk_length_s = gr.Number(label="Chunk Lengths (sec)", value=30, precision=0)
-                        nb_batch_size = gr.Number(label="Batch Size", value=24, precision=0)
+
+                    whisper_params, dd_file_format, cb_timestamp = self.create_whisper_parameters()
+
                     with gr.Row():
                         btn_run = gr.Button("GENERATE SUBTITLE FILE", variant="primary")
                     with gr.Row():
@@ -220,30 +231,6 @@ class App:
                         btn_openfolder = gr.Button('📂', scale=1)
 
                     params = [tb_youtubelink, dd_file_format, cb_timestamp]
-                    whisper_params = WhisperParameters(model_size=dd_model,
-                                                       lang=dd_lang,
-                                                       is_translate=cb_translate,
-                                                       beam_size=nb_beam_size,
-                                                       log_prob_threshold=nb_log_prob_threshold,
-                                                       no_speech_threshold=nb_no_speech_threshold,
-                                                       compute_type=dd_compute_type,
-                                                       best_of=nb_best_of,
-                                                       patience=nb_patience,
-                                                       condition_on_previous_text=cb_condition_on_previous_text,
-                                                       initial_prompt=tb_initial_prompt,
-                                                       temperature=sd_temperature,
-                                                       compression_ratio_threshold=nb_compression_ratio_threshold,
-                                                       vad_filter=cb_vad_filter,
-                                                       threshold=sd_threshold,
-                                                       min_speech_duration_ms=nb_min_speech_duration_ms,
-                                                       max_speech_duration_s=nb_max_speech_duration_s,
-                                                       min_silence_duration_ms=nb_min_silence_duration_ms,
-                                                       speech_pad_ms=nb_speech_pad_ms,
-                                                       chunk_length_s=nb_chunk_length_s,
-                                                       batch_size=nb_batch_size,
-                                                       is_diarize=cb_diarize,
-                                                       hf_token=tb_hf_token,
-                                                       diarization_device=dd_diarization_device)
 
                     btn_run.click(fn=self.whisper_inf.transcribe_youtube,
                                   inputs=params + whisper_params.as_list(),
@@ -251,48 +238,13 @@ class App:
                     tb_youtubelink.change(get_ytmetas, inputs=[tb_youtubelink],
                                           outputs=[img_thumbnail, tb_title, tb_description])
                     btn_openfolder.click(fn=lambda: self.open_folder("outputs"), inputs=None, outputs=None)
-                    dd_model.change(fn=self.on_change_models, inputs=[dd_model], outputs=[cb_translate])
 
                 with gr.TabItem("Mic"):  # tab3
                     with gr.Row():
                         mic_input = gr.Microphone(label="Record with Mic", type="filepath", interactive=True)
-                    with gr.Row():
-                        dd_model = gr.Dropdown(choices=self.whisper_inf.available_models, value="large-v2",
-                                               label="Model")
-                        dd_lang = gr.Dropdown(choices=["Automatic Detection"] + self.whisper_inf.available_langs,
-                                              value="Automatic Detection", label="Language")
-                        dd_file_format = gr.Dropdown(["SRT", "WebVTT", "txt"], value="SRT", label="File Format")
-                    with gr.Row():
-                        cb_translate = gr.Checkbox(value=False, label="Translate to English?", interactive=True)
-                    with gr.Accordion("Advanced Parameters", open=False):
-                        nb_beam_size = gr.Number(label="Beam Size", value=1, precision=0, interactive=True)
-                        nb_log_prob_threshold = gr.Number(label="Log Probability Threshold", value=-1.0, interactive=True)
-                        nb_no_speech_threshold = gr.Number(label="No Speech Threshold", value=0.6, interactive=True)
-                        dd_compute_type = gr.Dropdown(label="Compute Type", choices=self.whisper_inf.available_compute_types, value=self.whisper_inf.current_compute_type, interactive=True)
-                        nb_best_of = gr.Number(label="Best Of", value=5, interactive=True)
-                        nb_patience = gr.Number(label="Patience", value=1, interactive=True)
-                        cb_condition_on_previous_text = gr.Checkbox(label="Condition On Previous Text", value=True, interactive=True)
-                        tb_initial_prompt = gr.Textbox(label="Initial Prompt", value=None, interactive=True)
-                        sd_temperature = gr.Slider(label="Temperature", value=0, step=0.01, maximum=1.0, interactive=True)
-                    with gr.Accordion("VAD", open=False):
-                        cb_vad_filter = gr.Checkbox(label="Enable Silero VAD Filter", value=False, interactive=True)
-                        sd_threshold = gr.Slider(minimum=0.0, maximum=1.0, step=0.01, label="Speech Threshold", value=0.5, info="Lower it to be more sensitive to small sounds.")
-                        nb_min_speech_duration_ms = gr.Number(label="Minimum Speech Duration (ms)", precision=0, value=250)
-                        nb_max_speech_duration_s = gr.Number(label="Maximum Speech Duration (s)", value=9999)
-                        nb_min_silence_duration_ms = gr.Number(label="Minimum Silence Duration (ms)", precision=0, value=2000)
-                        nb_speech_pad_ms = gr.Number(label="Speech Padding (ms)", precision=0, value=400)
-                    with gr.Accordion("Diarization", open=False):
-                        cb_diarize = gr.Checkbox(label="Enable Diarization")
-                        tb_hf_token = gr.Text(label="HuggingFace Token", value="",
-                                              info="This is only needed the first time you download the model. If you already have models, you don't need to enter. "
-                                                   "To download the model, you must manually go to \"https://huggingface.co/pyannote/speaker-diarization-3.1\" and agree to their requirement.")
-                        dd_diarization_device = gr.Dropdown(label="Device",
-                                                            choices=self.whisper_inf.diarizer.get_available_device(),
-                                                            value=self.whisper_inf.diarizer.get_device())
-                    with gr.Accordion("Insanely Fast Whisper Parameters", open=False,
-                                      visible=isinstance(self.whisper_inf, InsanelyFastWhisperInference)):
-                        nb_chunk_length_s = gr.Number(label="Chunk Lengths (sec)", value=30, precision=0)
-                        nb_batch_size = gr.Number(label="Batch Size", value=24, precision=0)
+
+                    whisper_params, dd_file_format, cb_timestamp = self.create_whisper_parameters()
+
                     with gr.Row():
                         btn_run = gr.Button("GENERATE SUBTITLE FILE", variant="primary")
                     with gr.Row():
@@ -301,36 +253,11 @@ class App:
                         btn_openfolder = gr.Button('📂', scale=1)
 
                     params = [mic_input, dd_file_format]
-                    whisper_params = WhisperParameters(model_size=dd_model,
-                                                       lang=dd_lang,
-                                                       is_translate=cb_translate,
-                                                       beam_size=nb_beam_size,
-                                                       log_prob_threshold=nb_log_prob_threshold,
-                                                       no_speech_threshold=nb_no_speech_threshold,
-                                                       compute_type=dd_compute_type,
-                                                       best_of=nb_best_of,
-                                                       patience=nb_patience,
-                                                       condition_on_previous_text=cb_condition_on_previous_text,
-                                                       initial_prompt=tb_initial_prompt,
-                                                       temperature=sd_temperature,
-                                                       compression_ratio_threshold=nb_compression_ratio_threshold,
-                                                       vad_filter=cb_vad_filter,
-                                                       threshold=sd_threshold,
-                                                       min_speech_duration_ms=nb_min_speech_duration_ms,
-                                                       max_speech_duration_s=nb_max_speech_duration_s,
-                                                       min_silence_duration_ms=nb_min_silence_duration_ms,
-                                                       speech_pad_ms=nb_speech_pad_ms,
-                                                       chunk_length_s=nb_chunk_length_s,
-                                                       batch_size=nb_batch_size,
-                                                       is_diarize=cb_diarize,
-                                                       hf_token=tb_hf_token,
-                                                       diarization_device=dd_diarization_device)
 
                     btn_run.click(fn=self.whisper_inf.transcribe_mic,
                                   inputs=params + whisper_params.as_list(),
                                   outputs=[tb_indicator, files_subtitles])
                     btn_openfolder.click(fn=lambda: self.open_folder("outputs"), inputs=None, outputs=None)
-                    dd_model.change(fn=self.on_change_models, inputs=[dd_model], outputs=[cb_translate])
 
                 with gr.TabItem("T2T Translation"):  # tab 4
                     with gr.Row():
@@ -389,7 +316,8 @@ class App:
                             md_vram_table = gr.HTML(NLLB_VRAM_TABLE, elem_id="md_nllb_vram_table")
 
                     btn_run.click(fn=self.nllb_inf.translate_file,
-                                  inputs=[file_subs, dd_nllb_model, dd_nllb_sourcelang, dd_nllb_targetlang, nb_max_length, cb_timestamp],
+                                  inputs=[file_subs, dd_nllb_model, dd_nllb_sourcelang, dd_nllb_targetlang,
+                                          nb_max_length, cb_timestamp],
                                   outputs=[tb_indicator, files_subtitles])
 
                     btn_openfolder.click(fn=lambda: self.open_folder(os.path.join("outputs", "translations")),
@@ -412,10 +340,26 @@ class App:
 
         self.app.queue(api_open=False).launch(**launch_args)
 
+    @staticmethod
+    def open_folder(folder_path: str):
+        if os.path.exists(folder_path):
+            os.system(f"start {folder_path}")
+        else:
+            print(f"The folder {folder_path} does not exist.")
+
+    @staticmethod
+    def on_change_models(model_size: str):
+        translatable_model = ["large", "large-v1", "large-v2", "large-v3"]
+        if model_size not in translatable_model:
+            return gr.Checkbox(visible=False, value=False, interactive=False)
+        else:
+            return gr.Checkbox(visible=True, value=False, label="Translate to English?", interactive=True)
+
 
 # Create the parser for command-line arguments
 parser = argparse.ArgumentParser()
-parser.add_argument('--whisper_type', type=str, default="faster-whisper", help='A type of the whisper implementation between: ["whisper", "faster-whisper", "insanely-fast-whisper"]')
+parser.add_argument('--whisper_type', type=str, default="faster-whisper",
+                    help='A type of the whisper implementation between: ["whisper", "faster-whisper", "insanely-fast-whisper"]')
 parser.add_argument('--share', type=bool, default=False, nargs='?', const=True, help='Gradio share value')
 parser.add_argument('--server_name', type=str, default=None, help='Gradio server host')
 parser.add_argument('--server_port', type=int, default=None, help='Gradio server port')
@@ -425,11 +369,17 @@ parser.add_argument('--password', type=str, default=None, help='Gradio authentic
 parser.add_argument('--theme', type=str, default=None, help='Gradio Blocks theme')
 parser.add_argument('--colab', type=bool, default=False, nargs='?', const=True, help='Is colab user or not')
 parser.add_argument('--api_open', type=bool, default=False, nargs='?', const=True, help='enable api or not')
-parser.add_argument('--whisper_model_dir', type=str, default=os.path.join("models", "Whisper"), help='Directory path of the whisper model')
-parser.add_argument('--faster_whisper_model_dir', type=str, default=os.path.join("models", "Whisper", "faster-whisper"), help='Directory path of the faster-whisper model')
-parser.add_argument('--insanely_fast_whisper_model_dir', type=str, default=os.path.join("models", "Whisper", "insanely-fast-whisper"), help='Directory path of the insanely-fast-whisper model')
-parser.add_argument('--diarization_model_dir', type=str, default=os.path.join("models", "Diarization"), help='Directory path of the diarization model')
-parser.add_argument('--nllb_model_dir', type=str, default=os.path.join("models", "NLLB"), help='Directory path of the Facebook NLLB model')
+parser.add_argument('--whisper_model_dir', type=str, default=os.path.join("models", "Whisper"),
+                    help='Directory path of the whisper model')
+parser.add_argument('--faster_whisper_model_dir', type=str, default=os.path.join("models", "Whisper", "faster-whisper"),
+                    help='Directory path of the faster-whisper model')
+parser.add_argument('--insanely_fast_whisper_model_dir', type=str,
+                    default=os.path.join("models", "Whisper", "insanely-fast-whisper"),
+                    help='Directory path of the insanely-fast-whisper model')
+parser.add_argument('--diarization_model_dir', type=str, default=os.path.join("models", "Diarization"),
+                    help='Directory path of the diarization model')
+parser.add_argument('--nllb_model_dir', type=str, default=os.path.join("models", "NLLB"),
+                    help='Directory path of the Facebook NLLB model')
 parser.add_argument('--output_dir', type=str, default=os.path.join("outputs"), help='Directory path of the outputs')
 _args = parser.parse_args()
 

modules/whisper/faster_whisper_inference.py

@@ -5,6 +5,7 @@ import torch
 from typing import BinaryIO, Union, Tuple, List
 import faster_whisper
 from faster_whisper.vad import VadOptions
+import ast
 import ctranslate2
 import whisper
 import gradio as gr
@@ -62,6 +63,8 @@ class FasterWhisperInference(WhisperBase):
         if params.model_size != self.current_model_size or self.model is None or self.current_compute_type != params.compute_type:
             self.update_model(params.model_size, params.compute_type, progress)
 
+        params.suppress_tokens = self.format_suppress_tokens_str(params.suppress_tokens)
+
         segments, info = self.model.transcribe(
             audio=audio,
             language=params.lang,
@@ -73,6 +76,23 @@ class FasterWhisperInference(WhisperBase):
             patience=params.patience,
             temperature=params.temperature,
             compression_ratio_threshold=params.compression_ratio_threshold,
+            length_penalty=params.length_penalty,
+            repetition_penalty=params.repetition_penalty,
+            no_repeat_ngram_size=params.no_repeat_ngram_size,
+            prefix=params.prefix,
+            suppress_blank=params.suppress_blank,
+            suppress_tokens=params.suppress_tokens,
+            max_initial_timestamp=params.max_initial_timestamp,
+            word_timestamps=params.word_timestamps,
+            prepend_punctuations=params.prepend_punctuations,
+            append_punctuations=params.append_punctuations,
+            max_new_tokens=params.max_new_tokens,
+            chunk_length=params.chunk_length,
+            hallucination_silence_threshold=params.hallucination_silence_threshold,
+            hotwords=params.hotwords,
+            language_detection_threshold=params.language_detection_threshold,
+            language_detection_segments=params.language_detection_segments,
+            prompt_reset_on_temperature=params.prompt_reset_on_temperature
         )
         progress(0, desc="Loading audio..")
 
@@ -147,3 +167,13 @@ class FasterWhisperInference(WhisperBase):
             return "cuda"
         else:
             return "auto"
+
+    @staticmethod
+    def format_suppress_tokens_str(suppress_tokens_str: str) -> List[int]:
+        try:
+            suppress_tokens = ast.literal_eval(suppress_tokens_str)
+            if not isinstance(suppress_tokens, list) or not all(isinstance(item, int) for item in suppress_tokens):
+                raise ValueError("Invalid Suppress Tokens. The value must be type of List[int]")
+            return suppress_tokens
+        except Exception as e:
+            raise ValueError("Invalid Suppress Tokens. The value must be type of List[int]")

modules/whisper/whisper_parameter.py

@@ -15,6 +15,7 @@ class WhisperParameters:
     best_of: gr.Number
     patience: gr.Number
     condition_on_previous_text: gr.Checkbox
+    prompt_reset_on_temperature: gr.Slider
     initial_prompt: gr.Textbox
     temperature: gr.Slider
     compression_ratio_threshold: gr.Number
@@ -29,6 +30,22 @@ class WhisperParameters:
     is_diarize: gr.Checkbox
     hf_token: gr.Textbox
     diarization_device: gr.Dropdown
+    length_penalty: gr.Number
+    repetition_penalty: gr.Number
+    no_repeat_ngram_size: gr.Number
+    prefix: gr.Textbox
+    suppress_blank: gr.Checkbox
+    suppress_tokens: gr.Textbox
+    max_initial_timestamp: gr.Number
+    word_timestamps: gr.Checkbox
+    prepend_punctuations: gr.Textbox
+    append_punctuations: gr.Textbox
+    max_new_tokens: gr.Number
+    chunk_length: gr.Number
+    hallucination_silence_threshold: gr.Number
+    hotwords: gr.Textbox
+    language_detection_threshold: gr.Number
+    language_detection_segments: gr.Number
     """
     A data class for Gradio components of the Whisper Parameters. Use "before" Gradio pre-processing.
     This data class is used to mitigate the key-value problem between Gradio components and function parameters.
@@ -129,6 +146,62 @@ class WhisperParameters:
         
     diarization_device: gr.Dropdown
         This parameter is related with whisperx. Device to run diarization model
+        
+    length_penalty: 
+        This parameter is related to faster-whisper. Exponential length penalty constant.
+    
+    repetition_penalty: 
+        This parameter is related to faster-whisper. Penalty applied to the score of previously generated tokens
+        (set > 1 to penalize).
+
+    no_repeat_ngram_size:
+        This parameter is related to faster-whisper. Prevent repetitions of n-grams with this size (set 0 to disable).
+
+    prefix:
+        This parameter is related to faster-whisper. Optional text to provide as a prefix for the first window.
+
+    suppress_blank:
+        This parameter is related to faster-whisper. Suppress blank outputs at the beginning of the sampling.
+
+    suppress_tokens:
+        This parameter is related to faster-whisper. List of token IDs to suppress. -1 will suppress a default set
+        of symbols as defined in the model config.json file.
+
+    max_initial_timestamp:
+        This parameter is related to faster-whisper. The initial timestamp cannot be later than this.
+
+    word_timestamps:
+        This parameter is related to faster-whisper. Extract word-level timestamps using the cross-attention pattern
+        and dynamic time warping, and include the timestamps for each word in each segment.
+
+    prepend_punctuations:
+        This parameter is related to faster-whisper. If word_timestamps is True, merge these punctuation symbols
+        with the next word.
+
+    append_punctuations:
+        This parameter is related to faster-whisper. If word_timestamps is True, merge these punctuation symbols
+        with the previous word.
+
+    max_new_tokens:
+        This parameter is related to faster-whisper. Maximum number of new tokens to generate per-chunk. If not set,
+        the maximum will be set by the default max_length.
+
+    chunk_length:
+        This parameter is related to faster-whisper. The length of audio segments. If it is not None, it will overwrite the
+        default chunk_length of the FeatureExtractor.
+
+    hallucination_silence_threshold:
+        This parameter is related to faster-whisper. When word_timestamps is True, skip silent periods longer than this threshold
+        (in seconds) when a possible hallucination is detected.
+
+    hotwords:
+        This parameter is related to faster-whisper. Hotwords/hint phrases to provide the model with. Has no effect if prefix is not None.
+
+    language_detection_threshold:
+        This parameter is related to faster-whisper. If the maximum probability of the language tokens is higher than this value, the language is detected.
+
+    language_detection_segments:
+        This parameter is related to faster-whisper. Number of segments to consider for the language detection.
     """
 
     def as_list(self) -> list:
@@ -153,32 +226,7 @@ class WhisperParameters:
         WhisperValues
            Data class that has values of parameters
         """
-        return WhisperValues(
-            model_size=args[0],
-            lang=args[1],
-            is_translate=args[2],
-            beam_size=args[3],
-            log_prob_threshold=args[4],
-            no_speech_threshold=args[5],
-            compute_type=args[6],
-            best_of=args[7],
-            patience=args[8],
-            condition_on_previous_text=args[9],
-            initial_prompt=args[10],
-            temperature=args[11],
-            compression_ratio_threshold=args[12],
-            vad_filter=args[13],
-            threshold=args[14],
-            min_speech_duration_ms=args[15],
-            max_speech_duration_s=args[16],
-            min_silence_duration_ms=args[17],
-            speech_pad_ms=args[18],
-            chunk_length_s=args[19],
-            batch_size=args[20],
-            is_diarize=args[21],
-            hf_token=args[22],
-            diarization_device=args[23]
-        )
+        return WhisperValues(*args)
 
 
 @dataclass
@@ -193,6 +241,7 @@ class WhisperValues:
     best_of: int
     patience: float
     condition_on_previous_text: bool
+    prompt_reset_on_temperature: float
     initial_prompt: Optional[str]
     temperature: float
     compression_ratio_threshold: float
@@ -207,6 +256,22 @@ class WhisperValues:
     is_diarize: bool
     hf_token: str
     diarization_device: str
+    length_penalty: float
+    repetition_penalty: float
+    no_repeat_ngram_size: int
+    prefix: Optional[str]
+    suppress_blank: bool
+    suppress_tokens: Optional[str]
+    max_initial_timestamp: float
+    word_timestamps: bool
+    prepend_punctuations: Optional[str]
+    append_punctuations: Optional[str]
+    max_new_tokens: Optional[int]
+    chunk_length: Optional[int]
+    hallucination_silence_threshold: Optional[float]
+    hotwords: Optional[str]
+    language_detection_threshold: Optional[float]
+    language_detection_segments: int
     """
     A data class to use Whisper parameters.
     """