Add new Whisper models to models_directory.json and adjust transcription filter properties

data/models/models_directory.json

@@ -435,6 +435,116 @@
         }
       ]
     },
+    {
+      "friendly_name": "Whisper Base q8 (81Mb)",
+      "local_folder_name": "ggml-base-q8_0",
+      "type": "MODEL_TYPE_TRANSCRIPTION",
+      "files": [
+        {
+          "url": "https://huggingface.co/ggerganov/whisper.cpp/blob/main/ggml-base-q8_0.bin",
+          "sha256": ""
+        }
+      ]
+    },
+    {
+      "friendly_name": "Whisper Base English q8 (81Mb)",
+      "local_folder_name": "ggml-base-en-q8_0",
+      "type": "MODEL_TYPE_TRANSCRIPTION",
+      "files": [
+        {
+          "url": "https://huggingface.co/ggerganov/whisper.cpp/blob/main/ggml-base.en-q8_0.bin",
+          "sha256": ""
+        }
+      ]
+    },
+    {
+      "friendly_name": "Whisper Large v2 q8 (1.66Gb)",
+      "local_folder_name": "ggml-large-v2-q8_0",
+      "type": "MODEL_TYPE_TRANSCRIPTION",
+      "files": [
+        {
+          "url": "https://huggingface.co/ggerganov/whisper.cpp/blob/main/ggml-large-v2-q8_0.bin",
+          "sha256": ""
+        }
+      ]
+    },
+    {
+      "friendly_name": "Whisper Medium q8 (823Mb)",
+      "local_folder_name": "ggml-medium-q8_0",
+      "type": "MODEL_TYPE_TRANSCRIPTION",
+      "files": [
+        {
+          "url": "https://huggingface.co/ggerganov/whisper.cpp/blob/main/ggml-medium-q8_0.bin",
+          "sha256": ""
+        }
+      ]
+    },
+    {
+      "friendly_name": "Whisper Large v3 Trubo q8 (874Mb)",
+      "local_folder_name": "ggml-large-v3-turbo-q8_0",
+      "type": "MODEL_TYPE_TRANSCRIPTION",
+      "files": [
+        {
+          "url": "https://huggingface.co/ggerganov/whisper.cpp/blob/main/ggml-large-v3-turbo-q8_0.bin",
+          "sha256": ""
+        }
+      ]
+    },
+    {
+      "friendly_name": "Whisper Small q8 (264Mb)",
+      "local_folder_name": "ggml-small-q8_0",
+      "type": "MODEL_TYPE_TRANSCRIPTION",
+      "files": [
+        {
+          "url": "https://huggingface.co/ggerganov/whisper.cpp/blob/main/ggml-small-q8_0.bin",
+          "sha256": ""
+        }
+      ]
+    },
+    {
+      "friendly_name": "Whisper Medium English q8 (823Mb)",
+      "local_folder_name": "ggml-medium-en-q8_0",
+      "type": "MODEL_TYPE_TRANSCRIPTION",
+      "files": [
+        {
+          "url": "https://huggingface.co/ggerganov/whisper.cpp/blob/main/ggml-medium.en-q8_0.bin",
+          "sha256": ""
+        }
+      ]
+    },
+    {
+      "friendly_name": "Whisper Small English q8 (264Mb)",
+      "local_folder_name": "ggml-small-en-q8_0",
+      "type": "MODEL_TYPE_TRANSCRIPTION",
+      "files": [
+        {
+          "url": "https://huggingface.co/ggerganov/whisper.cpp/blob/main/ggml-small.en-q8_0.bin",
+          "sha256": ""
+        }
+      ]
+    },
+    {
+      "friendly_name": "Whisper Tiny q8 (43Mb)",
+      "local_folder_name": "ggml-tiny-q8_0",
+      "type": "MODEL_TYPE_TRANSCRIPTION",
+      "files": [
+        {
+          "url": "https://huggingface.co/ggerganov/whisper.cpp/blob/main/ggml-tiny-q8_0.bin",
+          "sha256": ""
+        }
+      ]
+    },
+    {
+      "friendly_name": "Whisper Tiny English q8 (43Mb)",
+      "local_folder_name": "ggml-tiny-en-q8_0",
+      "type": "MODEL_TYPE_TRANSCRIPTION",
+      "files": [
+        {
+          "url": "https://huggingface.co/ggerganov/whisper.cpp/blob/main/ggml-tiny.en-q8_0.bin",
+          "sha256": ""
+        }
+      ]
+    },
     {
       "friendly_name": "Whisper Base German Awareai (Marksdo)",
       "local_folder_name": "ggml-base-awareai.de",
@@ -4916,4 +5026,4 @@
       }
     }
   ]
-}
+}

src/transcription-filter-properties.cpp

@@ -626,7 +626,7 @@ void transcription_filter_defaults(obs_data_t *s)
 	obs_data_set_default_bool(s, "split_on_word", true);
 	obs_data_set_default_int(s, "max_tokens", 50);
 	obs_data_set_default_bool(s, "suppress_blank", false);
-	obs_data_set_default_bool(s, "suppress_non_speech_tokens", true);
+	obs_data_set_default_bool(s, "suppress_non_speech_tokens", false);
 	obs_data_set_default_double(s, "temperature", 0.1);
 	obs_data_set_default_double(s, "max_initial_ts", 1.0);
 	obs_data_set_default_double(s, "length_penalty", -1.0);

src/whisper-utils/vad-processing.cpp

@@ -24,9 +24,11 @@ int get_data_from_buf_and_resample(transcription_filter_data *gf,
 			return 1;
 		}
 
+#ifdef LOCALVOCAL_EXTRA_VERBOSE
 		obs_log(gf->log_level,
 			"segmentation: currently %lu bytes in the audio input buffer",
 			gf->input_buffers[0].size);
+#endif
 
 		// max number of frames is 10 seconds worth of audio
 		const size_t max_num_frames = gf->sample_rate * 10;
@@ -76,8 +78,10 @@ int get_data_from_buf_and_resample(transcription_filter_data *gf,
 		}
 	}
 
+#ifdef LOCALVOCAL_EXTRA_VERBOSE
 	obs_log(gf->log_level, "found %d frames from info buffer.", num_frames_from_infos);
 	gf->last_num_frames = num_frames_from_infos;
+#endif
 
 	{
 		// resample to 16kHz
@@ -95,11 +99,13 @@ int get_data_from_buf_and_resample(transcription_filter_data *gf,
 
 		circlebuf_push_back(&gf->resampled_buffer, resampled_16khz[0],
 				    resampled_16khz_frames * sizeof(float));
+#ifdef LOCALVOCAL_EXTRA_VERBOSE
 		obs_log(gf->log_level,
 			"resampled: %d channels, %d frames, %f ms, current size: %lu bytes",
 			(int)gf->channels, (int)resampled_16khz_frames,
 			(float)resampled_16khz_frames / WHISPER_SAMPLE_RATE * 1000.0f,
 			gf->resampled_buffer.size);
+#endif
 	}
 
 	return 0;
@@ -129,8 +135,10 @@ vad_state vad_based_segmentation(transcription_filter_data *gf, vad_state last_v
 	circlebuf_pop_front(&gf->resampled_buffer, vad_input.data(),
 			    vad_input.size() * sizeof(float));
 
+#ifdef LOCALVOCAL_EXTRA_VERBOSE
 	obs_log(gf->log_level, "sending %d frames to vad, %d windows, reset state? %s",
 		vad_input.size(), vad_num_windows, (!last_vad_state.vad_on) ? "yes" : "no");
+#endif
 	{
 		ProfileScope("vad->process");
 		gf->vad->process(vad_input, !last_vad_state.vad_on);
@@ -144,7 +152,9 @@ vad_state vad_based_segmentation(transcription_filter_data *gf, vad_state last_v
 
 	std::vector<timestamp_t> stamps = gf->vad->get_speech_timestamps();
 	if (stamps.size() == 0) {
+#ifdef LOCALVOCAL_EXTRA_VERBOSE
 		obs_log(gf->log_level, "VAD detected no speech in %u frames", vad_input.size());
+#endif
 		if (last_vad_state.vad_on) {
 			obs_log(gf->log_level, "Last VAD was ON: segment end -> send to inference");
 			run_inference_and_callbacks(gf, last_vad_state.start_ts_offest_ms,