added multilingual support for whisper

Author: LostRuins

expose.h

@@ -192,6 +192,7 @@ struct whisper_generation_inputs
     const char * prompt = nullptr;
     const char * audio_data = nullptr;
     const bool suppress_non_speech = false;
+    const char * langcode = nullptr;
     const bool quiet = false;
 };
 struct whisper_generation_outputs

kcpp_docs.embd

@@ -1344,13 +1344,27 @@
                             "application/json": {
                                "example": {
                                   "prompt": "",
+                                  "suppress_non_speech" : false,
+                                  "langcode": "en",
                                   "audio_data": "base64_wav_data",
                                },
                                "schema": {
                                   "properties": {
                                      "audio_data": {
                                         "type": "string",
                                         "description": "Base64 respresentation of a 16-bit 16kHz wave file to be transcribed to text."
+                                     },
+                                     "prompt": {
+                                        "type": "string",
+                                        "description": "Prompt to steer the transcription."
+                                     },
+                                     "langcode": {
+                                        "type": "string",
+                                        "description": "Two letter language code, or use auto to autodetect."
+                                     },
+                                     "suppress_non_speech": {
+                                        "type": "boolean",
+                                        "description": "Prevent noise tokens, always generate words for speech."
                                      }
                                   },
                                   "type": "object"

klite.embd

@@ -12,7 +12,7 @@ Current version indicated by LITEVER below.
 -->
 
 <script>
-	const LITEVER = 201;
+	const LITEVER = 202;
 	const urlParams = new URLSearchParams(window.location.search);
 	var localflag = true;
 	const STORAGE_PREFIX = (localflag?"e_":"")+"kaihordewebui_";
@@ -3044,6 +3044,7 @@ Current version indicated by LITEVER below.
 		narrate_only_dialog: false,
 		voice_end_delay: 300,
 		voice_suppress_nonspeech: false,
+		voice_langcode: "auto",
 		tts_speed: 1.0,
 		image_styles: "",
 		image_negprompt: "",
@@ -10330,6 +10331,7 @@ initializeInstructUIFunctionality();
 		document.getElementById("tts_speed").value = localsettings.tts_speed;
 		document.getElementById("voice_end_delay").value = localsettings.voice_end_delay;
 		document.getElementById("voice_suppress_nonspeech").checked = localsettings.voice_suppress_nonspeech;
+		document.getElementById("voice_langcode").value = localsettings.voice_langcode;
 		toggle_opmode();
 
 		//sd models display
@@ -10618,6 +10620,7 @@ initializeInstructUIFunctionality();
 		localsettings.tts_speed = document.getElementById("tts_speed").value;
 		localsettings.voice_end_delay = document.getElementById("voice_end_delay").value;
 		localsettings.voice_suppress_nonspeech = (document.getElementById("voice_suppress_nonspeech").checked?true:false);
+		localsettings.voice_langcode = document.getElementById("voice_langcode").value;
 		localsettings.auto_ctxlen = (document.getElementById("auto_ctxlen").checked ? true : false);
 		localsettings.auto_genamt = (document.getElementById("auto_genamt").checked ? true : false);
 
@@ -15299,7 +15302,8 @@ initializeInstructUIFunctionality();
 		let payload = {
 			"audio_data": dataurl,
 			"prompt": "",
-			"suppress_non_speech": localsettings.voice_suppress_nonspeech
+			"suppress_non_speech": localsettings.voice_suppress_nonspeech,
+			"langcode": localsettings.voice_langcode,
 		};
 		fetch(apply_proxy_url(custom_kobold_endpoint + koboldcpp_transcribe_endpoint), {
 			method: 'POST',
@@ -15349,7 +15353,8 @@ initializeInstructUIFunctionality();
 							let payload = {
 								"audio_data": dataurl,
 								"prompt": "",
-								"suppress_non_speech": (document.getElementById("voice_suppress_nonspeech").checked?true:false)
+								"suppress_non_speech": (document.getElementById("voice_suppress_nonspeech").checked?true:false),
+								"langcode": document.getElementById("voice_langcode").value
 							};
 							fetch(apply_proxy_url(custom_kobold_endpoint + koboldcpp_transcribe_endpoint), {
 								method: 'POST',
@@ -20229,6 +20234,12 @@ initializeInstructUIFunctionality();
 							<div class="justifyleft" style="padding:2px"  title="Suppress non-speech (e.g. music and sounds) from transcription">Suppress Non-Speech </div>
 						  	<input title="Suppress Non-Speech" type="checkbox" id="voice_suppress_nonspeech" style="margin:0px 0px 0px auto;">
 						</div>
+						<div class="inlinelabel" style="font-size: 11px;">
+							<div class="justifyleft" style="padding:2px"  title="Language Code">Language </div>
+							<input class="settinglabel miniinput" type="text" placeholder="en" value="auto" id="voice_langcode" style="height:18px; width: 36px; padding: 2px;">
+
+						</div>
+
 						<div class="inlinelabel" style="font-size: 11px;">
 							<div class="justifyleft" style="padding:3px">Voice Delay: </div>
 							<input title="Voice Delay Milliseconds" type="text" inputmode="decimal" value="300" id="voice_end_delay" style="width:40px">

koboldcpp.py

@@ -273,6 +273,7 @@ class whisper_generation_inputs(ctypes.Structure):
     _fields_ = [("prompt", ctypes.c_char_p),
                 ("audio_data", ctypes.c_char_p),
                 ("suppress_non_speech", ctypes.c_bool),
+                ("langcode", ctypes.c_char_p),
                 ("quiet", ctypes.c_bool)]
 
 class whisper_generation_outputs(ctypes.Structure):
@@ -1252,6 +1253,9 @@ def whisper_generate(genparams):
     inputs.prompt = prompt.encode("UTF-8")
     inputs.audio_data = audio_data.encode("UTF-8")
     inputs.quiet = is_quiet
+    lc = genparams.get("langcode", "auto")
+    lc = lc.strip().lower() if (lc and lc.strip().lower()!="") else "auto"
+    inputs.langcode = lc.encode("UTF-8")
     inputs.suppress_non_speech = genparams.get("suppress_non_speech", False)
     ret = handle.whisper_generate(inputs)
     outstr = ""

otherarch/whispercpp/whisper.cpp

@@ -5355,13 +5355,19 @@ int whisper_full_with_state(
 
         const auto lang_id = whisper_lang_auto_detect_with_state(ctx, state, 0, params.n_threads, probs.data());
         if (lang_id < 0) {
-            WHISPER_LOG_ERROR("%s: failed to auto-detect language\n", __func__);
+            if(params.debug_mode)
+            {
+                printf("\n%s: failed to auto-detect language\n", __func__);
+            }
             return -3;
         }
         state->lang_id = lang_id;
         params.language = whisper_lang_str(lang_id);
 
-        WHISPER_LOG_INFO("%s: auto-detected language: %s (p = %f)\n", __func__, params.language, probs[whisper_lang_id(params.language)]);
+        if(params.debug_mode)
+        {
+            printf("\n%s: auto-detected language: %s (p = %f)\n", __func__, params.language, probs[whisper_lang_id(params.language)]);
+        }
         if (params.detect_language) {
             return 0;
         }
@@ -5477,7 +5483,11 @@ int whisper_full_with_state(
     std::vector<whisper_token> prompt_init = { whisper_token_sot(ctx), };
 
     if (whisper_is_multilingual(ctx)) {
-        const int lang_id = whisper_lang_id(params.language);
+        int lang_id = whisper_lang_id(params.language);
+        if(lang_id<0)
+        {
+            lang_id = 0; //default to english
+        }
         state->lang_id = lang_id;
         prompt_init.push_back(whisper_token_lang(ctx, lang_id));
         if (params.translate) {

otherarch/whispercpp/whisper_adapter.cpp

@@ -217,6 +217,7 @@ whisper_generation_outputs whispertype_generate(const whisper_generation_inputs
 
     const std::string b64data = std::string(inputs.audio_data);
     const std::string initprompt = std::string(inputs.prompt);
+    const std::string langcode = std::string(inputs.langcode);
 
     std::vector<float> pcmf32;               // mono-channel F32 PCM
     std::vector<std::vector<float>> pcmf32s; // stereo-channel F32 PCM
@@ -236,7 +237,7 @@ whisper_generation_outputs whispertype_generate(const whisper_generation_inputs
     wparams.print_timestamps = false;
     wparams.print_special    = false;
     wparams.translate        = false;
-    wparams.language         = "auto";
+    wparams.language         = langcode.c_str();
     wparams.detect_language  = false;
     wparams.n_threads        = 4;
     wparams.n_max_text_ctx   = wparams.n_max_text_ctx;
@@ -248,7 +249,7 @@ whisper_generation_outputs whispertype_generate(const whisper_generation_inputs
     wparams.split_on_word    = false;
     wparams.audio_ctx        = 0;
     wparams.speed_up         = false;
-    wparams.debug_mode       = false;
+    wparams.debug_mode       = (whisperdebugmode==1);
     wparams.tdrz_enable      = false;
     wparams.suppress_regex   = nullptr;
     wparams.suppress_non_speech_tokens = inputs.suppress_non_speech;