Better TTS with autoplay

- Adds "still_streaming" to shared module for extensions to know if generation is complete
- Changed TTS extension with new options:
   - Show text under the audio widget
   - Automatically play the audio once text generation finishes
   - manage the generated wav files (only keep files for finished generations, optional max file limit)
   - [wip] ability to change voice pitch and speed
- added 'tensorboard' to requirements, since python sent "tensorboard not found" errors after a fresh installation.

extensions/silero_tts/requirements.txt

@@ -4,3 +4,4 @@ pydub
 PyYAML
 torch
 torchaudio
+simpleaudio

extensions/silero_tts/script.py

@@ -4,20 +4,36 @@ from pathlib import Path
 import gradio as gr
 import torch
 
+import modules.shared as shared
+import simpleaudio as sa
+
 torch._C._jit_set_profiling_mode(False)
 
 params = {
     'activate': True,
-    'speaker': 'en_56',
+    'speaker': 'en_5',
     'language': 'en',
     'model_id': 'v3_en',
     'sample_rate': 48000,
     'device': 'cpu',
+    'max_wavs': 20,
+    'play_audio': True,
+    'show_text': True,
 }
 current_params = params.copy()
 voices_by_gender = ['en_99', 'en_45', 'en_18', 'en_117', 'en_49', 'en_51', 'en_68', 'en_0', 'en_26', 'en_56', 'en_74', 'en_5', 'en_38', 'en_53', 'en_21', 'en_37', 'en_107', 'en_10', 'en_82', 'en_16', 'en_41', 'en_12', 'en_67', 'en_61', 'en_14', 'en_11', 'en_39', 'en_52', 'en_24', 'en_97', 'en_28', 'en_72', 'en_94', 'en_36', 'en_4', 'en_43', 'en_88', 'en_25', 'en_65', 'en_6', 'en_44', 'en_75', 'en_91', 'en_60', 'en_109', 'en_85', 'en_101', 'en_108', 'en_50', 'en_96', 'en_64', 'en_92', 'en_76', 'en_33', 'en_116', 'en_48', 'en_98', 'en_86', 'en_62', 'en_54', 'en_95', 'en_55', 'en_111', 'en_3', 'en_83', 'en_8', 'en_47', 'en_59', 'en_1', 'en_2', 'en_7', 'en_9', 'en_13', 'en_15', 'en_17', 'en_19', 'en_20', 'en_22', 'en_23', 'en_27', 'en_29', 'en_30', 'en_31', 'en_32', 'en_34', 'en_35', 'en_40', 'en_42', 'en_46', 'en_57', 'en_58', 'en_63', 'en_66', 'en_69', 'en_70', 'en_71', 'en_73', 'en_77', 'en_78', 'en_79', 'en_80', 'en_81', 'en_84', 'en_87', 'en_89', 'en_90', 'en_93', 'en_100', 'en_102', 'en_103', 'en_104', 'en_105', 'en_106', 'en_110', 'en_112', 'en_113', 'en_114', 'en_115']
 wav_idx = 0
 
+table = str.maketrans({
+    "<": "&lt;",
+    ">": "&gt;",
+    "&": "&amp;",
+    "'": "&apos;",
+    '"': "&quot;",
+})
+def xmlesc(txt):
+    return txt.translate(table)
+
 def load_model():
     model, example_text = torch.hub.load(repo_or_dir='snakers4/silero-models', model='silero_tts', language=params['language'], speaker=params['model_id'])
     model.to(params['device'])
@@ -58,20 +74,45 @@ def output_modifier(string):
     if params['activate'] == False:
         return string
 
+    orig_string = string
     string = remove_surrounded_chars(string)
     string = string.replace('"', '')
     string = string.replace('“', '')
     string = string.replace('\n', ' ')
     string = string.strip()
 
+    auto_playable=True
     if string == '':
-        string = 'empty reply, try regenerating'
-
+            string = 'empty reply, try regenerating'
+            auto_playable=False
+            
+
+    #x-slow, slow, medium, fast, x-fast
+    #x-low, low, medium, high, x-high
+    #prosody='<prosody rate="fast" pitch="medium">'
+    prosody='<prosody rate="fast">'
+    string ='<speak>'+prosody+xmlesc(string)+'</prosody></speak>'
+        
     output_file = Path(f'extensions/silero_tts/outputs/{wav_idx:06d}.wav')
-    audio = model.save_wav(text=string, speaker=params['speaker'], sample_rate=int(params['sample_rate']), audio_path=str(output_file))
-
+    audio = model.save_wav(ssml_text=string, speaker=params['speaker'], sample_rate=int(params['sample_rate']), audio_path=str(output_file))
     string = f'<audio src="file/{output_file.as_posix()}" controls></audio>'
-    wav_idx += 1
+    
+    #reset if too many wavs. set max to -1 for unlimited.
+    if wav_idx < params['max_wavs'] and params['max_wavs'] > 0:
+        #only increment if starting a new stream, else replace during streaming. Does not update duration on webui sometimes?
+        if not shared.still_streaming:
+            wav_idx += 1
+    else:
+        wav_idx = 0
+        
+    if params['show_text']:
+        string+='\n\n'+orig_string
+    
+    #if params['play_audio'] == True and auto_playable and shared.stop_everything:
+    if params['play_audio'] == True and auto_playable and not shared.still_streaming:
+        stop_autoplay()
+        wave_obj = sa.WaveObject.from_wave_file(output_file.as_posix())
+        wave_obj.play()
 
     return string
 
@@ -84,11 +125,20 @@ def bot_prefix_modifier(string):
 
     return string
 
+def stop_autoplay():
+    sa.stop_all()
+
 def ui():
     # Gradio elements
     activate = gr.Checkbox(value=params['activate'], label='Activate TTS')
+    show_text = gr.Checkbox(value=params['show_text'], label='Show message text under audio player')
+    play_audio = gr.Checkbox(value=params['play_audio'], label='Play TTS automatically')
+    stop_audio = gr.Button("Stop Auto-Play")
     voice = gr.Dropdown(value=params['speaker'], choices=voices_by_gender, label='TTS voice')
 
     # Event functions to update the parameters in the backend
     activate.change(lambda x: params.update({"activate": x}), activate, None)
+    play_audio.change(lambda x: params.update({"play_audio": x}), play_audio, None)
+    show_text.change(lambda x: params.update({"show_text": x}), show_text, None)
+    stop_audio.click(stop_autoplay)
     voice.change(lambda x: params.update({"speaker": x}), voice, None)

modules/shared.py

@@ -12,6 +12,7 @@ is_LLaMA = False
 history = {'internal': [], 'visible': []}
 character = 'None'
 stop_everything = False
+still_streaming = False
 
 # UI elements (buttons, sliders, HTML, etc)
 gradio = {}

modules/text_generation.py

@@ -182,6 +182,7 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi
     # Generate the reply 8 tokens at a time
     else:
         yield formatted_outputs(original_question, shared.model_name)
+        shared.still_streaming = True
         for i in tqdm(range(max_new_tokens//8+1)):
             with torch.no_grad():
                 output = eval(f"shared.model.generate({', '.join(generate_params)}){cuda}")[0]
@@ -191,8 +192,7 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi
             reply = decode(output)
             if not (shared.args.chat or shared.args.cai_chat):
                 reply = original_question + apply_extensions(reply[len(question):], "output")
-            yield formatted_outputs(reply, shared.model_name)
-
+            
             if not shared.args.flexgen:
                 if output[-1] == n:
                     break
@@ -201,6 +201,13 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi
                 if np.count_nonzero(input_ids[0] == n) < np.count_nonzero(output == n):
                     break
                 input_ids = np.reshape(output, (1, output.shape[0]))
+                
+            #Mid-stream yield, ran if no breaks
+            yield formatted_outputs(reply, shared.model_name)
 
             if shared.soft_prompt:
                 inputs_embeds, filler_input_ids = generate_softprompt_input_tensors(input_ids)
+                
+        #Stream finished from max tokens or break. Do final yield.
+        shared.still_streaming = False
+        yield formatted_outputs(reply, shared.model_name)

requirements.txt

@@ -6,3 +6,4 @@ numpy
 rwkv==0.0.6
 safetensors==0.2.8
 git+https://github.com/huggingface/transformers
+tensorboard