Chore: panic is not solved (WIP)

Makefile

@@ -1,4 +1,4 @@
-.PHONY: setconfig run lint lintall install-linters setup-whisper build-whisper download-whisper-model docker-up docker-down docker-logs noextra-run installdelve checkdelve fetch-onnx install-onnx-deps
+.PHONY: setconfig run lint lintall install-linters setup-whisper build-whisper download-whisper-model docker-up docker-down docker-logs noextra-run installdelve checkdelve fetch-onnx install-onnx-deps fetch-kokoro-voices install-espeak
 
 run: setconfig
 	go build -tags extra -o gf-lt && ./gf-lt
@@ -33,6 +33,9 @@ lintall: lint
 fetch-onnx:
 	mkdir -p onnx/embedgemma && curl -o onnx/embedgemma/config.json -L https://huggingface.co/onnx-community/embeddinggemma-300m-ONNX/resolve/main/config.json && curl -o onnx/embedgemma/tokenizer.json -L https://huggingface.co/onnx-community/embeddinggemma-300m-ONNX/resolve/main/tokenizer.json && curl -o onnx/embedgemma/model_q4.onnx -L https://huggingface.co/onnx-community/embeddinggemma-300m-ONNX/resolve/main/onnx/model_q4.onnx && curl -o onnx/embedgemma/model_q4.onnx_data -L https://huggingface.co/onnx-community/embeddinggemma-300m-ONNX/resolve/main/onnx/model_q4.onnx_data?download=true
 
+fetch-kokoro-onnx:
+	mkdir -p onnx/kokoro && curl -o onnx/kokoro/config.json -L https://huggingface.co/onnx-community/Kokoro-82M-v1.0-ONNX/resolve/main/config.json && curl -o onnx/kokoro/tokenizer.json -L https://huggingface.co/onnx-community/Kokoro-82M-v1.0-ONNX/resolve/main/tokenizer.json && curl -o onnx/kokoro/model_quantized.onnx -L https://huggingface.co/onnx-community/Kokoro-82M-v1.0-ONNX/resolve/main/onnx/model_quantized.onnx && curl -o onnx/kokoro/voices.bin -L https://github.com/thewh1teagle/kokoro-onnx/releases/download/model-files-v1.0/voices-v1.0.bin
+
 install-onnx-deps: ## Install ONNX Runtime with CUDA support (or CPU fallback)
 	@echo "=== ONNX Runtime Installer ===" && \
 	echo "" && \
@@ -194,3 +197,25 @@ docker-logs-whisper: ## View logs from Whisper STT service only
 docker-logs-kokoro: ## View logs from Kokoro TTS service only
 	@echo "Displaying logs from Kokoro TTS service..."
 	docker-compose -f batteries/docker-compose.yml logs -f kokoro-tts
+
+# Kokoro ONNX TTS Setup
+install-espeak: ## Install espeak-ng for phoneme tokenization
+	@echo "=== Installing espeak-ng ===" && \
+	if command -v espeak-ng >/dev/null 2>&1; then \
+		echo "espeak-ng is already installed:" && \
+		espeak-ng --version && \
+		exit 0; \
+	fi && \
+	echo "Installing espeak-ng..." && \
+	sudo apt-get update && \
+	sudo apt-get install -y espeak-ng espeak && \
+	echo "espeak-ng installed successfully!" && \
+	espeak-ng --version
+
+fetch-kokoro-voices: ## Download Kokoro voice files (PyTorch format)
+	@echo "=== Downloading Kokoro voices ===" && \
+	mkdir -p onnx/kokoro/voices && \
+	echo "Downloading af_bella voice..." && \
+	curl -L -o onnx/kokoro/voices/af_bella.pt https://raw.githubusercontent.com/hexgrad/kokoro/main/kokoro/voices/af_heart.pt && \
+	echo "Voice file downloaded to onnx/kokoro/voices/" && \
+	ls -lh onnx/kokoro/voices/

bot.go

@@ -418,7 +418,9 @@ func fetchLCPModelsWithStatus() (*models.LCPModels, error) {
 	if err := json.NewDecoder(resp.Body).Decode(data); err != nil {
 		return nil, err
 	}
+	localModelsMu.Lock()
 	localModelsData = data
+	localModelsMu.Unlock()
 	return data, nil
 }
 
@@ -1497,7 +1499,7 @@ func init() {
 	// load cards
 	basicCard.Role = cfg.AssistantRole
 	logLevel.Set(slog.LevelInfo)
-	logger = slog.New(slog.NewTextHandler(logfile, &slog.HandlerOptions{Level: logLevel}))
+	logger = slog.New(slog.NewTextHandler(logfile, &slog.HandlerOptions{Level: logLevel, AddSource: true}))
 	store = storage.NewProviderSQL(cfg.DBPATH, logger)
 	if store == nil {
 		cancel()

config/config.go

@@ -61,6 +61,10 @@ type Config struct {
 	TTS_SPEED    float32 `toml:"TTS_SPEED"`
 	TTS_PROVIDER string  `toml:"TTS_PROVIDER"`
 	TTS_LANGUAGE string  `toml:"TTS_LANGUAGE"`
+	// Kokoro ONNX TTS
+	KokoroModelPath  string `toml:"KokoroModelPath"`
+	KokoroVoicesPath string `toml:"KokoroVoicesPath"`
+	KokoroVoice      string `toml:"KokoroVoice"`
 	// STT
 	STT_TYPE          string `toml:"STT_TYPE"` // WHISPER_SERVER, WHISPER_BINARY
 	STT_URL           string `toml:"STT_URL"`

extra/google_tts.go

@@ -1,218 +0,0 @@
-//go:build extra
-// +build extra
-
-package extra
-
-import (
-	"fmt"
-	"gf-lt/models"
-	"io"
-	"log/slog"
-	"os/exec"
-	"strings"
-	"sync"
-
-	google_translate_tts "github.com/GrailFinder/google-translate-tts"
-	"github.com/neurosnap/sentences/english"
-)
-
-type GoogleTranslateOrator struct {
-	logger *slog.Logger
-	mu     sync.Mutex
-	speech *google_translate_tts.Speech
-	// fields for playback control
-	cmd    *exec.Cmd
-	cmdMu  sync.Mutex
-	stopCh chan struct{}
-	// text buffer and interrupt flag
-	textBuffer strings.Builder
-	interrupt  bool
-	Speed      float32
-}
-
-func (o *GoogleTranslateOrator) stoproutine() {
-	for {
-		<-TTSDoneChan
-		o.logger.Debug("orator got done signal")
-		o.Stop()
-		for len(TTSTextChan) > 0 {
-			<-TTSTextChan
-		}
-		o.mu.Lock()
-		o.textBuffer.Reset()
-		o.interrupt = true
-		o.mu.Unlock()
-	}
-}
-
-func (o *GoogleTranslateOrator) readroutine() {
-	tokenizer, _ := english.NewSentenceTokenizer(nil)
-	for {
-		select {
-		case chunk := <-TTSTextChan:
-			o.mu.Lock()
-			o.interrupt = false
-			_, err := o.textBuffer.WriteString(chunk)
-			if err != nil {
-				o.logger.Warn("failed to write to stringbuilder", "error", err)
-				o.mu.Unlock()
-				continue
-			}
-			text := o.textBuffer.String()
-			sentences := tokenizer.Tokenize(text)
-			o.logger.Debug("adding chunk", "chunk", chunk, "text", text, "sen-len", len(sentences))
-			if len(sentences) <= 1 {
-				o.mu.Unlock()
-				continue
-			}
-			completeSentences := sentences[:len(sentences)-1]
-			remaining := sentences[len(sentences)-1].Text
-			o.textBuffer.Reset()
-			o.textBuffer.WriteString(remaining)
-			o.mu.Unlock()
-			for _, sentence := range completeSentences {
-				o.mu.Lock()
-				interrupted := o.interrupt
-				o.mu.Unlock()
-				if interrupted {
-					return
-				}
-				cleanedText := models.CleanText(sentence.Text)
-				if cleanedText == "" {
-					continue
-				}
-				o.logger.Debug("calling Speak with sentence", "sent", cleanedText)
-				if err := o.Speak(cleanedText); err != nil {
-					o.logger.Error("tts failed", "sentence", cleanedText, "error", err)
-				}
-			}
-		case <-TTSFlushChan:
-			o.logger.Debug("got flushchan signal start")
-			// lln is done get the whole message out
-			if len(TTSTextChan) > 0 { // otherwise might get stuck
-				for chunk := range TTSTextChan {
-					o.mu.Lock()
-					_, err := o.textBuffer.WriteString(chunk)
-					o.mu.Unlock()
-					if err != nil {
-						o.logger.Warn("failed to write to stringbuilder", "error", err)
-						continue
-					}
-					if len(TTSTextChan) == 0 {
-						break
-					}
-				}
-			}
-			o.mu.Lock()
-			remaining := o.textBuffer.String()
-			remaining = models.CleanText(remaining)
-			o.textBuffer.Reset()
-			o.mu.Unlock()
-			if remaining == "" {
-				continue
-			}
-			o.logger.Debug("calling Speak with remainder", "rem", remaining)
-			sentencesRem := tokenizer.Tokenize(remaining)
-			for _, rs := range sentencesRem { // to avoid dumping large volume of text
-				o.mu.Lock()
-				interrupt := o.interrupt
-				o.mu.Unlock()
-				if interrupt {
-					break
-				}
-				if err := o.Speak(rs.Text); err != nil {
-					o.logger.Error("tts failed", "sentence", rs.Text, "error", err)
-				}
-			}
-		}
-	}
-}
-
-func (o *GoogleTranslateOrator) GetLogger() *slog.Logger {
-	return o.logger
-}
-
-func (o *GoogleTranslateOrator) Speak(text string) error {
-	o.logger.Debug("fn: Speak is called", "text-len", len(text))
-	// Generate MP3 data directly as an io.Reader
-	reader, err := o.speech.GenerateSpeech(text)
-	if err != nil {
-		return fmt.Errorf("generate speech failed: %w", err)
-	}
-	// Wrap in io.NopCloser since GenerateSpeech returns io.Reader (no close needed)
-	body := io.NopCloser(reader)
-	defer body.Close()
-	// Build ffplay command with optional speed filter
-	args := []string{"-nodisp", "-autoexit"}
-	if o.Speed > 0.1 && o.Speed != 1.0 {
-		// atempo range is 0.5 to 2.0; you might clamp it here
-		args = append(args, "-af", fmt.Sprintf("atempo=%.2f", o.Speed))
-	}
-	args = append(args, "-i", "pipe:0")
-	cmd := exec.Command("ffplay", args...)
-	stdin, err := cmd.StdinPipe()
-	if err != nil {
-		return fmt.Errorf("failed to get stdin pipe: %w", err)
-	}
-	o.cmdMu.Lock()
-	o.cmd = cmd
-	o.stopCh = make(chan struct{})
-	o.cmdMu.Unlock()
-	if err := cmd.Start(); err != nil {
-		return fmt.Errorf("failed to start ffplay: %w", err)
-	}
-	copyErr := make(chan error, 1)
-	go func() {
-		_, err := io.Copy(stdin, body)
-		stdin.Close()
-		copyErr <- err
-	}()
-	done := make(chan error, 1)
-	go func() {
-		done <- cmd.Wait()
-	}()
-	select {
-	case <-o.stopCh:
-		if o.cmd != nil && o.cmd.Process != nil {
-			o.cmd.Process.Kill()
-		}
-		<-done
-		return nil
-	case copyErrVal := <-copyErr:
-		if copyErrVal != nil {
-			if o.cmd != nil && o.cmd.Process != nil {
-				o.cmd.Process.Kill()
-			}
-			<-done
-			return copyErrVal
-		}
-		return <-done
-	case err := <-done:
-		return err
-	}
-}
-
-func (o *GoogleTranslateOrator) Stop() {
-	o.cmdMu.Lock()
-	defer o.cmdMu.Unlock()
-	// Signal any running Speak to stop
-	if o.stopCh != nil {
-		select {
-		case <-o.stopCh: // already closed
-		default:
-			close(o.stopCh)
-		}
-		o.stopCh = nil
-	}
-	// Kill the external player process if it's still running
-	if o.cmd != nil && o.cmd.Process != nil {
-		o.cmd.Process.Kill()
-		o.cmd.Wait() // clean up zombie process
-		o.cmd = nil
-	}
-	// Also reset text buffer and interrupt flag (with o.mu)
-	o.mu.Lock()
-	o.textBuffer.Reset()
-	o.interrupt = true
-	o.mu.Unlock()
-}

extra/kokoro.go

@@ -1,259 +0,0 @@
-//go:build extra
-// +build extra
-
-package extra
-
-import (
-	"bytes"
-	"encoding/json"
-	"fmt"
-	"gf-lt/models"
-	"io"
-	"log/slog"
-	"net/http"
-	"os/exec"
-	"strings"
-	"sync"
-
-	"github.com/neurosnap/sentences/english"
-)
-
-type KokoroOrator struct {
-	logger   *slog.Logger
-	mu       sync.Mutex
-	URL      string
-	Format   models.AudioFormat
-	Stream   bool
-	Speed    float32
-	Language string
-	Voice    string
-	// fields for playback control
-	cmd    *exec.Cmd
-	cmdMu  sync.Mutex
-	stopCh chan struct{}
-	// textBuffer, interrupt etc. remain the same
-	textBuffer strings.Builder
-	interrupt  bool
-}
-
-func (o *KokoroOrator) GetLogger() *slog.Logger {
-	return o.logger
-}
-
-func (o *KokoroOrator) Speak(text string) error {
-	o.logger.Debug("fn: Speak is called", "text-len", len(text))
-	body, err := o.requestSound(text)
-	if err != nil {
-		return fmt.Errorf("request failed: %w", err)
-	}
-	defer body.Close()
-	cmd := exec.Command("ffplay", "-nodisp", "-autoexit", "-i", "pipe:0")
-	stdin, err := cmd.StdinPipe()
-	if err != nil {
-		return fmt.Errorf("failed to get stdin pipe: %w", err)
-	}
-	o.cmdMu.Lock()
-	o.cmd = cmd
-	o.stopCh = make(chan struct{})
-	o.cmdMu.Unlock()
-	if err := cmd.Start(); err != nil {
-		return fmt.Errorf("failed to start ffplay: %w", err)
-	}
-	// Copy audio in background
-	copyErr := make(chan error, 1)
-	go func() {
-		_, err := io.Copy(stdin, body)
-		stdin.Close()
-		copyErr <- err
-	}()
-	// Wait for player in background
-	done := make(chan error, 1)
-	go func() {
-		done <- cmd.Wait()
-	}()
-	// Wait for BOTH copy and player, but ensure we block until done
-	select {
-	case <-o.stopCh:
-		// Stop requested: kill player and wait for it to exit
-		if o.cmd != nil && o.cmd.Process != nil {
-			o.cmd.Process.Kill()
-		}
-		<-done // Wait for process to actually exit
-		return nil
-	case copyErrVal := <-copyErr:
-		if copyErrVal != nil {
-			// Copy failed: kill player and wait
-			if o.cmd != nil && o.cmd.Process != nil {
-				o.cmd.Process.Kill()
-			}
-			<-done
-			return copyErrVal
-		}
-		// Copy succeeded, now wait for playback to complete
-		return <-done
-	case err := <-done:
-		// Playback finished normally (copy must have succeeded or player would have exited early)
-		return err
-	}
-}
-func (o *KokoroOrator) requestSound(text string) (io.ReadCloser, error) {
-	if o.URL == "" {
-		return nil, fmt.Errorf("TTS URL is empty")
-	}
-	payload := map[string]interface{}{
-		"input":           text,
-		"voice":           o.Voice,
-		"response_format": o.Format,
-		"download_format": o.Format,
-		"stream":          o.Stream,
-		"speed":           o.Speed,
-		// "return_download_link": true,
-		"lang_code": o.Language,
-	}
-	payloadBytes, err := json.Marshal(payload)
-	if err != nil {
-		return nil, fmt.Errorf("failed to marshal payload: %w", err)
-	}
-	req, err := http.NewRequest("POST", o.URL, bytes.NewBuffer(payloadBytes)) //nolint:noctx
-	if err != nil {
-		return nil, fmt.Errorf("failed to create request: %w", err)
-	}
-	req.Header.Set("accept", "application/json")
-	req.Header.Set("Content-Type", "application/json")
-	resp, err := http.DefaultClient.Do(req)
-	if err != nil {
-		return nil, fmt.Errorf("request failed: %w", err)
-	}
-	if resp.StatusCode != http.StatusOK {
-		defer resp.Body.Close()
-		return nil, fmt.Errorf("unexpected status code: %d", resp.StatusCode)
-	}
-	return resp.Body, nil
-}
-
-func (o *KokoroOrator) stoproutine() {
-	for {
-		<-TTSDoneChan
-		o.logger.Debug("orator got done signal")
-		// 1. Stop any ongoing playback (kills external player, closes stopCh)
-		o.Stop()
-		// 2. Drain any pending text chunks
-		for len(TTSTextChan) > 0 {
-			<-TTSTextChan
-		}
-		// 3. Reset internal state
-		o.mu.Lock()
-		o.textBuffer.Reset()
-		o.interrupt = true
-		o.mu.Unlock()
-	}
-}
-
-func (o *KokoroOrator) Stop() {
-	o.cmdMu.Lock()
-	defer o.cmdMu.Unlock()
-	// Signal any running Speak to stop
-	if o.stopCh != nil {
-		select {
-		case <-o.stopCh: // already closed
-		default:
-			close(o.stopCh)
-		}
-		o.stopCh = nil
-	}
-	// Kill the external player process if it's still running
-	if o.cmd != nil && o.cmd.Process != nil {
-		o.cmd.Process.Kill()
-		o.cmd.Wait() // clean up zombie process
-		o.cmd = nil
-	}
-	// Also reset text buffer and interrupt flag (with o.mu)
-	o.mu.Lock()
-	o.textBuffer.Reset()
-	o.interrupt = true
-	o.mu.Unlock()
-}
-
-func (o *KokoroOrator) readroutine() {
-	tokenizer, _ := english.NewSentenceTokenizer(nil)
-	for {
-		select {
-		case chunk := <-TTSTextChan:
-			o.mu.Lock()
-			o.interrupt = false
-			_, err := o.textBuffer.WriteString(chunk)
-			if err != nil {
-				o.logger.Warn("failed to write to stringbuilder", "error", err)
-				o.mu.Unlock()
-				continue
-			}
-			text := o.textBuffer.String()
-			sentences := tokenizer.Tokenize(text)
-			o.logger.Debug("adding chunk", "chunk", chunk, "text", text, "sen-len", len(sentences))
-			if len(sentences) <= 1 {
-				o.mu.Unlock()
-				continue
-			}
-			completeSentences := sentences[:len(sentences)-1]
-			remaining := sentences[len(sentences)-1].Text
-			o.textBuffer.Reset()
-			o.textBuffer.WriteString(remaining)
-			o.mu.Unlock()
-			for _, sentence := range completeSentences {
-				o.mu.Lock()
-				interrupted := o.interrupt
-				o.mu.Unlock()
-				if interrupted {
-					return
-				}
-				cleanedText := models.CleanText(sentence.Text)
-				if cleanedText == "" {
-					continue
-				}
-				o.logger.Debug("calling Speak with sentence", "sent", cleanedText)
-				if err := o.Speak(cleanedText); err != nil {
-					o.logger.Error("tts failed", "sentence", cleanedText, "error", err)
-				}
-			}
-		case <-TTSFlushChan:
-			o.logger.Debug("got flushchan signal start")
-			// lln is done get the whole message out
-			if len(TTSTextChan) > 0 { // otherwise might get stuck
-				for chunk := range TTSTextChan {
-					o.mu.Lock()
-					_, err := o.textBuffer.WriteString(chunk)
-					o.mu.Unlock()
-					if err != nil {
-						o.logger.Warn("failed to write to stringbuilder", "error", err)
-						continue
-					}
-					if len(TTSTextChan) == 0 {
-						break
-					}
-				}
-			}
-			// flush remaining text
-			o.mu.Lock()
-			remaining := o.textBuffer.String()
-			remaining = models.CleanText(remaining)
-			o.textBuffer.Reset()
-			o.mu.Unlock()
-			if remaining == "" {
-				continue
-			}
-			o.logger.Debug("calling Speak with remainder", "rem", remaining)
-			sentencesRem := tokenizer.Tokenize(remaining)
-			for _, rs := range sentencesRem { // to avoid dumping large volume of text
-				o.mu.Lock()
-				interrupt := o.interrupt
-				o.mu.Unlock()
-				if interrupt {
-					break
-				}
-				if err := o.Speak(rs.Text); err != nil {
-					o.logger.Error("tts failed", "sentence", rs, "error", err)
-				}
-			}
-		}
-	}
-}

extra/kokoro_onnx.go

@@ -0,0 +1,421 @@
+//go:build extra
+// +build extra
+
+package extra
+
+import (
+	"bytes"
+	"fmt"
+	"gf-lt/models"
+	"gf-lt/onnx"
+	"log/slog"
+	"os/exec"
+	"strings"
+	"sync"
+	"time"
+
+	"github.com/gopxl/beep/v2"
+	"github.com/gopxl/beep/v2/speaker"
+	"github.com/gopxl/beep/v2/wav"
+	"github.com/neurosnap/sentences/english"
+	"github.com/yalue/onnxruntime_go"
+)
+
+// KokoroONNXOrator implements Kokoro TTS using ONNX runtime
+type KokoroONNXOrator struct {
+	logger        *slog.Logger
+	mu            sync.Mutex
+	session       *onnxruntime_go.DynamicAdvancedSession
+	phonemeMap    map[string]int
+	espeakCmd     string
+	voice         string
+	speed         float32
+	styleVector   []float32
+	currentStream *beep.Ctrl
+	currentDone   chan bool
+	textBuffer    strings.Builder
+	interrupt     bool
+	modelLoaded   bool
+	modelPath     string
+	voicesPath    string
+}
+
+// Phoneme to token ID mapping from Kokoro tokenizer.json
+var kokoroPhonemeMap = map[string]int{
+	"$": 0, ";": 1, ":": 2, ",": 3, ".": 4, "!": 5, "?": 6, "—": 9, "…": 10, "\"": 11, "(": 12, ")": 13, "“": 14, "”": 15, " ": 16, "̃": 17, "ˢ": 18, "ˤ": 19, "˦": 20, "˨": 21, "ᾝ": 22, "⭧": 23,
+	"A": 24, "I": 25, "O": 31, "Q": 33, "S": 35, "T": 36, "W": 39, "Y": 41, "ʲ": 42,
+	"a": 43, "b": 44, "c": 45, "d": 46, "e": 47, "f": 48, "h": 50, "i": 51, "j": 52, "k": 53, "l": 54, "m": 55, "n": 56, "o": 57, "p": 58, "q": 59, "r": 60, "s": 61, "t": 62, "u": 63, "v": 64, "w": 65, "x": 66, "y": 67, "z": 68,
+	"ɑ": 69, "ɐ": 70, "ɒ": 71, "æ": 72, "β": 75, "ɔ": 76, "ɕ": 77, "ç": 78, "ɖ": 80, "ð": 81, "˔": 82, "ə": 83, "ɚ": 85, "ɛ": 86, "ɜ": 87, "ɟ": 90, "ɡ": 92, "ɥ": 99, "ɨ": 101, "ɪ": 102, "ɝ": 103, "ɯ": 110, "ɰ": 111, "ŋ": 112, "ɳ": 113, "ɲ": 114, "ɴ": 115, "ø": 116, "ɸ": 118, "θ": 119, "œ": 120, "ɹ": 123, "ɾ": 125, "ɺ": 126, "ʁ": 128, "ɽ": 129, "ʂ": 130, "ʃ": 131, "ʈ": 132, "˧": 133, "ʊ": 135, "ʋ": 136, "ʌ": 138, "ɢ": 139, "ɣ": 140, "χ": 142, "ʎ": 143, "ʒ": 147, "ʔ": 148,
+	"ˈ": 156, "ˌ": 157, "ː": 158, "̰": 162, "̊": 164, "↕": 169, "→": 171, "↗": 172, "↘": 173, "ᶻ": 177,
+}
+
+func (o *KokoroONNXOrator) ensureInitialized(modelPath string) error {
+	if o.modelLoaded {
+		return nil
+	}
+	o.mu.Lock()
+	defer o.mu.Unlock()
+	if o.modelLoaded {
+		return nil
+	}
+	if modelPath == "" {
+		o.logger.Error("modelPath is empty, cannot load ONNX model")
+		return fmt.Errorf("modelPath is empty, set KokoroModelPath in config")
+	}
+	// Initialize ONNX runtime (shared with embedder)
+	if err := onnx.Init(); err != nil {
+		o.logger.Error("ONNX init failed", "error", err)
+		return fmt.Errorf("ONNX init failed: %w", err)
+	}
+	if onnx.HasCUDASupport() {
+		o.logger.Info("ONNX using CUDA")
+	} else {
+		o.logger.Info("ONNX using CPU fallback")
+	}
+	if o.phonemeMap == nil {
+		o.phonemeMap = kokoroPhonemeMap
+	}
+	if o.espeakCmd == "" {
+		o.espeakCmd = "espeak-ng"
+		if _, err := exec.LookPath(o.espeakCmd); err != nil {
+			o.espeakCmd = "espeak"
+			if _, err := exec.LookPath(o.espeakCmd); err != nil {
+				return fmt.Errorf("espeak-ng or espeak not found. Install with: sudo apt-get install espeak-ng")
+			}
+		}
+	}
+	o.logger.Info("using espeak command", "cmd", o.espeakCmd)
+	// Load voice embedding if not already loaded
+	if o.styleVector == nil {
+		voiceName := o.voice
+		if voiceName == "" {
+			voiceName = "af_bella"
+		}
+		if o.voicesPath != "" {
+			styleVec, err := onnx.LoadVoice(o.voicesPath, voiceName)
+			if err != nil {
+				o.logger.Warn("failed to load voice, using zeros", "error", err, "voice", voiceName)
+				o.styleVector = make([]float32, 256)
+			} else {
+				// Shape is (510, 1, 256), we want the last 256 values (or first? let's use mean or just pick one)
+				// Actually, let's average across all 510 to get a single 256-dim vector
+				if len(styleVec) != 510*256 {
+					o.logger.Error("voice embedding has unexpected size", "len", len(styleVec))
+					err = fmt.Errorf("voice embedding has unexpected size", "len", len(styleVec))
+					return err
+				}
+				o.styleVector = make([]float32, 256)
+				for i := 0; i < 256; i++ {
+					var sum float32
+					for j := 0; j < 510; j++ {
+						sum += styleVec[j*256+i]
+					}
+					o.styleVector[i] = sum / 510.0
+				}
+				o.logger.Info("loaded voice embedding", "voice", voiceName)
+			}
+		} else {
+			o.logger.Warn("no voices path configured, using zeros for style")
+			o.styleVector = make([]float32, 256)
+		}
+	}
+	opts, err := onnx.NewSessionOptions()
+	if err != nil {
+		return fmt.Errorf("failed to create session options: %w", err)
+	}
+	defer func() { _ = opts.Destroy() }()
+	if onnx.HasCUDASupport() {
+		o.logger.Info("session options created with CUDA")
+	} else {
+		o.logger.Info("session options created with CPU")
+	}
+	session, err := onnxruntime_go.NewDynamicAdvancedSession(
+		modelPath,
+		[]string{"input_ids", "style", "speed"},
+		[]string{"waveform"},
+		opts,
+	)
+	if err != nil {
+		o.logger.Error("failed to create ONNX session", "error", err)
+		return fmt.Errorf("failed to create ONNX session: %w", err)
+	}
+	o.session = session
+	o.modelLoaded = true
+	o.logger.Info("Kokoro ONNX model loaded successfully", "model", modelPath)
+	return nil
+}
+
+func (o *KokoroONNXOrator) textToPhonemes(text string) (string, error) {
+	cmd := exec.Command(o.espeakCmd, "-x", "-q", text)
+	output, err := cmd.Output()
+	if err != nil {
+		o.logger.Error("espeak failed", "error", err, "cmd", o.espeakCmd, "text", text)
+		return "", fmt.Errorf("espeak failed: %w", err)
+	}
+
+	phonemeStr := strings.TrimSpace(string(output))
+	return phonemeStr, nil
+}
+
+func (o *KokoroONNXOrator) phonemesToTokens(phonemeStr string) ([]int, error) {
+	if phonemeStr == "" {
+		o.logger.Error("empty phoneme string")
+		return nil, fmt.Errorf("empty phoneme string")
+	}
+	// Iterate over each character in the phoneme string
+	tokens := make([]int, 0)
+	for _, ch := range phonemeStr {
+		chStr := string(ch)
+		if tokenID, ok := o.phonemeMap[chStr]; ok {
+			tokens = append(tokens, tokenID)
+		}
+	}
+	if len(tokens) == 0 {
+		o.logger.Error("no phonemes mapped to tokens", "phonemeStr", phonemeStr)
+		return nil, fmt.Errorf("no valid phonemes mapped to tokens")
+	}
+	return tokens, nil
+}
+
+func (o *KokoroONNXOrator) generateAudio(text string) ([]float32, error) {
+	if err := o.ensureInitialized(o.modelPath); err != nil {
+		o.logger.Error("ensureInitialized failed", "error", err)
+		return nil, err
+	}
+	phonemeStr, err := o.textToPhonemes(text)
+	if err != nil {
+		o.logger.Error("phoneme conversion failed", "error", err)
+		return nil, fmt.Errorf("phoneme conversion failed: %w", err)
+	}
+	tokens, err := o.phonemesToTokens(phonemeStr)
+	if err != nil {
+		o.logger.Error("token conversion failed", "error", err)
+		return nil, fmt.Errorf("token conversion failed: %w", err)
+	}
+	if len(tokens) > 510 {
+		return nil, fmt.Errorf("text too long: %d tokens (max 510)", len(tokens))
+	}
+	tokens = append([]int{0}, tokens...)
+	tokens = append(tokens, 0)
+	inputIDs := make([]int64, len(tokens))
+	for i, t := range tokens {
+		inputIDs[i] = int64(t)
+	}
+	inputTensor, err := onnxruntime_go.NewTensor[int64](
+		onnxruntime_go.NewShape(1, int64(len(inputIDs))),
+		inputIDs,
+	)
+	if err != nil {
+		o.logger.Error("failed to create input tensor", "error", err)
+		return nil, fmt.Errorf("failed to create input tensor: %w", err)
+	}
+	defer func() { _ = inputTensor.Destroy() }()
+	styleTensor, err := onnxruntime_go.NewTensor[float32](
+		onnxruntime_go.NewShape(1, 256),
+		o.styleVector,
+	)
+	if err != nil {
+		o.logger.Error("failed to create style tensor", "error", err)
+		return nil, fmt.Errorf("failed to create style tensor: %w", err)
+	}
+	defer func() { _ = styleTensor.Destroy() }()
+	speedTensor, err := onnxruntime_go.NewTensor[float32](
+		onnxruntime_go.NewShape(1),
+		[]float32{o.speed},
+	)
+	if err != nil {
+		o.logger.Error("failed to create speed tensor", "error", err)
+		return nil, fmt.Errorf("failed to create speed tensor: %w", err)
+	}
+	defer func() { _ = speedTensor.Destroy() }()
+	outputTensor, err := onnxruntime_go.NewEmptyTensor[float32](
+		onnxruntime_go.NewShape(1, 512),
+	)
+	if err != nil {
+		o.logger.Error("failed to create output tensor", "error", err)
+		return nil, fmt.Errorf("failed to create output tensor: %w", err)
+	}
+	defer func() { _ = outputTensor.Destroy() }()
+	err = o.session.Run(
+		[]onnxruntime_go.Value{inputTensor, styleTensor, speedTensor},
+		[]onnxruntime_go.Value{outputTensor},
+	)
+	if err != nil {
+		o.logger.Error("ONNX inference failed", "error", err)
+		return nil, fmt.Errorf("ONNX inference failed: %w", err)
+	}
+	audioData := outputTensor.GetData()
+	if len(audioData) == 0 {
+		o.logger.Error("empty audio output from ONNX")
+		return nil, fmt.Errorf("empty audio output")
+	}
+	audio := make([]float32, len(audioData))
+	copy(audio, audioData)
+	return audio, nil
+}
+
+func (o *KokoroONNXOrator) Speak(text string) error {
+	audio, err := o.generateAudio(text)
+	if err != nil {
+		o.logger.Error("audio generation failed", "error", err)
+		return fmt.Errorf("audio generation failed: %w", err)
+	}
+	// Create streamer for encoding
+	encodeStreamer := beep.StreamerFunc(func(samples [][2]float64) (n int, ok bool) {
+		for i := range samples {
+			if i >= len(audio) {
+				return i, false
+			}
+			samples[i][0] = float64(audio[i])
+			samples[i][1] = float64(audio[i])
+		}
+		return len(audio), true
+	})
+	buf := &seekableBuffer{new(bytes.Buffer)}
+	err = wav.Encode(buf, encodeStreamer, beep.Format{
+		SampleRate:  24000,
+		NumChannels: 1,
+		Precision:   2,
+	})
+	if err != nil {
+		o.logger.Error("wav encoding failed", "error", err)
+		return fmt.Errorf("wav encoding failed: %w", err)
+	}
+	decodedStreamer, format, err := wav.Decode(bytes.NewReader(buf.Bytes()))
+	if err != nil {
+		o.logger.Error("wav decode failed", "error", err)
+		return fmt.Errorf("wav decode failed: %w", err)
+	}
+	defer decodedStreamer.Close()
+	if err := speaker.Init(format.SampleRate, format.SampleRate.N(time.Second/10)); err != nil {
+		o.logger.Error("speaker init failed", "error", err)
+		return fmt.Errorf("speaker init failed: %w", err)
+	}
+	o.logger.Info("playing audio", "sampleRate", format.SampleRate, "channels", format.NumChannels)
+	done := make(chan bool)
+	o.mu.Lock()
+	o.currentDone = done
+	o.currentStream = &beep.Ctrl{Streamer: beep.Seq(decodedStreamer, beep.Callback(func() {
+		o.mu.Lock()
+		close(done)
+		o.currentStream = nil
+		o.currentDone = nil
+		o.mu.Unlock()
+	})), Paused: false}
+	o.mu.Unlock()
+	speaker.Play(o.currentStream)
+	<-done
+	return nil
+}
+
+func (o *KokoroONNXOrator) Stop() {
+	speaker.Lock()
+	defer speaker.Unlock()
+	o.mu.Lock()
+	defer o.mu.Unlock()
+	if o.currentStream != nil {
+		o.currentStream.Streamer = nil
+	}
+}
+
+func (o *KokoroONNXOrator) GetLogger() *slog.Logger {
+	return o.logger
+}
+
+func (o *KokoroONNXOrator) stoproutine() {
+	for {
+		<-TTSDoneChan
+		o.Stop()
+		for len(TTSTextChan) > 0 {
+			<-TTSTextChan
+		}
+		o.mu.Lock()
+		o.textBuffer.Reset()
+		if o.currentDone != nil {
+			select {
+			case o.currentDone <- true:
+			default:
+			}
+		}
+		o.interrupt = true
+		o.mu.Unlock()
+	}
+}
+
+func (o *KokoroONNXOrator) readroutine() {
+	tokenizer, _ := english.NewSentenceTokenizer(nil)
+	for {
+		select {
+		case chunk := <-TTSTextChan:
+			o.mu.Lock()
+			o.interrupt = false
+			_, err := o.textBuffer.WriteString(chunk)
+			if err != nil {
+				o.logger.Warn("failed to write to buffer", "error", err)
+				o.mu.Unlock()
+				continue
+			}
+			text := o.textBuffer.String()
+			sentences := tokenizer.Tokenize(text)
+			if len(sentences) <= 1 {
+				o.mu.Unlock()
+				continue
+			}
+			completeSentences := sentences[:len(sentences)-1]
+			remaining := sentences[len(sentences)-1].Text
+			o.textBuffer.Reset()
+			o.textBuffer.WriteString(remaining)
+			o.mu.Unlock()
+			for _, sentence := range completeSentences {
+				o.mu.Lock()
+				interrupted := o.interrupt
+				o.mu.Unlock()
+				if interrupted {
+					return
+				}
+				cleanedText := models.CleanText(sentence.Text)
+				if cleanedText == "" {
+					continue
+				}
+				o.logger.Info("KokoroONNX speak", "text", cleanedText)
+				if err := o.Speak(cleanedText); err != nil {
+					o.logger.Error("KokoroONNX tts failed", "text", cleanedText, "error", err)
+				}
+			}
+		case <-TTSFlushChan:
+			if len(TTSTextChan) > 0 {
+				for chunk := range TTSTextChan {
+					o.mu.Lock()
+					_, err := o.textBuffer.WriteString(chunk)
+					o.mu.Unlock()
+					if err != nil {
+						continue
+					}
+					if len(TTSTextChan) == 0 {
+						break
+					}
+				}
+			}
+			o.mu.Lock()
+			remaining := o.textBuffer.String()
+			remaining = models.CleanText(remaining)
+			o.textBuffer.Reset()
+			o.mu.Unlock()
+			if remaining == "" {
+				continue
+			}
+			sentencesRem := tokenizer.Tokenize(remaining)
+			for _, rs := range sentencesRem {
+				o.mu.Lock()
+				interrupt := o.interrupt
+				o.mu.Unlock()
+				if interrupt {
+					break
+				}
+				if err := o.Speak(rs.Text); err != nil {
+					o.logger.Error("tts failed", "text", rs.Text, "error", err)
+				}
+			}
+		}
+	}
+}

extra/tts.go

@@ -4,13 +4,25 @@
 package extra
 
 import (
+	"bytes"
+	"encoding/json"
+	"fmt"
 	"gf-lt/config"
 	"gf-lt/models"
+	"io"
 	"log/slog"
+	"net/http"
 	"os"
 	"strings"
+	"sync"
+	"time"
 
 	google_translate_tts "github.com/GrailFinder/google-translate-tts"
+	"github.com/GrailFinder/google-translate-tts/handlers"
+	"github.com/gopxl/beep/v2"
+	"github.com/gopxl/beep/v2/mp3"
+	"github.com/gopxl/beep/v2/speaker"
+	"github.com/neurosnap/sentences/english"
 )
 
 var (
@@ -20,6 +32,14 @@ var (
 	// endsWithPunctuation = regexp.MustCompile(`[;.!?]$`)
 )
 
+type seekableBuffer struct {
+	*bytes.Buffer
+}
+
+func (s *seekableBuffer) Seek(offset int64, whence int) (int64, error) {
+	return 0, nil
+}
+
 type Orator interface {
 	Speak(text string) error
 	Stop()
@@ -27,6 +47,142 @@ type Orator interface {
 	GetLogger() *slog.Logger
 }
 
+// impl https://github.com/remsky/Kokoro-FastAPI
+type KokoroOrator struct {
+	logger        *slog.Logger
+	mu            sync.Mutex
+	URL           string
+	Format        models.AudioFormat
+	Stream        bool
+	Speed         float32
+	Language      string
+	Voice         string
+	currentStream *beep.Ctrl // Added for playback control
+	currentDone   chan bool
+	textBuffer    strings.Builder
+	interrupt     bool
+	// textBuffer bytes.Buffer
+}
+
+// Google Translate TTS implementation
+type GoogleTranslateOrator struct {
+	logger        *slog.Logger
+	mu            sync.Mutex
+	speech        *google_translate_tts.Speech
+	currentStream *beep.Ctrl
+	currentDone   chan bool
+	textBuffer    strings.Builder
+	interrupt     bool
+}
+
+func (o *KokoroOrator) stoproutine() {
+	for {
+		<-TTSDoneChan
+		o.logger.Debug("orator got done signal")
+		o.Stop()
+		// drain the channel
+		for len(TTSTextChan) > 0 {
+			<-TTSTextChan
+		}
+		o.mu.Lock()
+		o.textBuffer.Reset()
+		if o.currentDone != nil {
+			select {
+			case o.currentDone <- true:
+			default:
+				// Channel might be closed, ignore
+			}
+		}
+		o.interrupt = true
+		o.mu.Unlock()
+	}
+}
+
+func (o *KokoroOrator) readroutine() {
+	tokenizer, _ := english.NewSentenceTokenizer(nil)
+	for {
+		select {
+		case chunk := <-TTSTextChan:
+			o.mu.Lock()
+			o.interrupt = false
+			_, err := o.textBuffer.WriteString(chunk)
+			if err != nil {
+				o.logger.Warn("failed to write to stringbuilder", "error", err)
+				o.mu.Unlock()
+				continue
+			}
+			text := o.textBuffer.String()
+			sentences := tokenizer.Tokenize(text)
+			o.logger.Debug("adding chunk", "chunk", chunk, "text", text, "sen-len", len(sentences))
+			if len(sentences) <= 1 {
+				o.mu.Unlock()
+				continue
+			}
+			completeSentences := sentences[:len(sentences)-1]
+			remaining := sentences[len(sentences)-1].Text
+			o.textBuffer.Reset()
+			o.textBuffer.WriteString(remaining)
+			o.mu.Unlock()
+
+			for _, sentence := range completeSentences {
+				o.mu.Lock()
+				interrupted := o.interrupt
+				o.mu.Unlock()
+				if interrupted {
+					return
+				}
+				cleanedText := models.CleanText(sentence.Text)
+				if cleanedText == "" {
+					continue
+				}
+				o.logger.Debug("calling Speak with sentence", "sent", cleanedText)
+				if err := o.Speak(cleanedText); err != nil {
+					o.logger.Error("tts failed", "sentence", cleanedText, "error", err)
+				}
+			}
+		case <-TTSFlushChan:
+			o.logger.Debug("got flushchan signal start")
+			// lln is done get the whole message out
+			if len(TTSTextChan) > 0 { // otherwise might get stuck
+				for chunk := range TTSTextChan {
+					o.mu.Lock()
+					_, err := o.textBuffer.WriteString(chunk)
+					o.mu.Unlock()
+					if err != nil {
+						o.logger.Warn("failed to write to stringbuilder", "error", err)
+						continue
+					}
+					if len(TTSTextChan) == 0 {
+						break
+					}
+				}
+			}
+			// flush remaining text
+			o.mu.Lock()
+			remaining := o.textBuffer.String()
+			remaining = models.CleanText(remaining)
+			o.textBuffer.Reset()
+			o.mu.Unlock()
+			if remaining == "" {
+				continue
+			}
+			o.logger.Debug("calling Speak with remainder", "rem", remaining)
+			sentencesRem := tokenizer.Tokenize(remaining)
+			for _, rs := range sentencesRem { // to avoid dumping large volume of text
+				o.mu.Lock()
+				interrupt := o.interrupt
+				o.mu.Unlock()
+				if interrupt {
+					break
+				}
+				if err := o.Speak(rs.Text); err != nil {
+					o.logger.Error("tts failed", "sentence", rs, "error", err)
+				}
+			}
+		}
+	}
+}
+
 func NewOrator(log *slog.Logger, cfg *config.Config) Orator {
 	provider := cfg.TTS_PROVIDER
 	if provider == "" {
@@ -46,6 +202,18 @@ func NewOrator(log *slog.Logger, cfg *config.Config) Orator {
 		go orator.readroutine()
 		go orator.stoproutine()
 		return orator
+	case "kokoro_onnx":
+		log.Info("Initializing Kokoro ONNX TTS", "modelPath", cfg.KokoroModelPath, "voicesPath", cfg.KokoroVoicesPath, "voice", cfg.KokoroVoice, "speed", cfg.TTS_SPEED)
+		orator := &KokoroONNXOrator{
+			logger:     log,
+			modelPath:  cfg.KokoroModelPath,
+			voicesPath: cfg.KokoroVoicesPath,
+			speed:      cfg.TTS_SPEED,
+			voice:      cfg.KokoroVoice,
+		}
+		go orator.readroutine()
+		go orator.stoproutine()
+		return orator
 	default:
 		language := cfg.TTS_LANGUAGE
 		if language == "" {
@@ -56,14 +224,270 @@ func NewOrator(log *slog.Logger, cfg *config.Config) Orator {
 			Language: language,
 			Proxy:    "", // Proxy not supported
 			Speed:    cfg.TTS_SPEED,
+			Handler:  &handlers.Beep{},
 		}
 		orator := &GoogleTranslateOrator{
 			logger: log,
 			speech: speech,
-			Speed:  cfg.TTS_SPEED,
 		}
 		go orator.readroutine()
 		go orator.stoproutine()
 		return orator
 	}
 }
+
+func (o *KokoroOrator) GetLogger() *slog.Logger {
+	return o.logger
+}
+
+func (o *KokoroOrator) requestSound(text string) (io.ReadCloser, error) {
+	if o.URL == "" {
+		return nil, fmt.Errorf("TTS URL is empty")
+	}
+	payload := map[string]interface{}{
+		"input":           text,
+		"voice":           o.Voice,
+		"response_format": o.Format,
+		"download_format": o.Format,
+		"stream":          o.Stream,
+		"speed":           o.Speed,
+		// "return_download_link": true,
+		"lang_code": o.Language,
+	}
+	payloadBytes, err := json.Marshal(payload)
+	if err != nil {
+		return nil, fmt.Errorf("failed to marshal payload: %w", err)
+	}
+	req, err := http.NewRequest("POST", o.URL, bytes.NewBuffer(payloadBytes)) //nolint:noctx
+	if err != nil {
+		return nil, fmt.Errorf("failed to create request: %w", err)
+	}
+	req.Header.Set("accept", "application/json")
+	req.Header.Set("Content-Type", "application/json")
+	resp, err := http.DefaultClient.Do(req)
+	if err != nil {
+		return nil, fmt.Errorf("request failed: %w", err)
+	}
+	if resp.StatusCode != http.StatusOK {
+		defer resp.Body.Close()
+		return nil, fmt.Errorf("unexpected status code: %d", resp.StatusCode)
+	}
+	return resp.Body, nil
+}
+
+func (o *KokoroOrator) Speak(text string) error {
+	o.logger.Debug("fn: Speak is called", "text-len", len(text))
+	body, err := o.requestSound(text)
+	if err != nil {
+		o.logger.Error("request failed", "error", err)
+		return fmt.Errorf("request failed: %w", err)
+	}
+	defer body.Close()
+	// Decode the mp3 audio from response body
+	streamer, format, err := mp3.Decode(body)
+	if err != nil {
+		o.logger.Error("mp3 decode failed", "error", err)
+		return fmt.Errorf("mp3 decode failed: %w", err)
+	}
+	defer streamer.Close()
+	// here it spams with errors that speaker cannot be initialized more than once, but how would we deal with many audio records then?
+	if err := speaker.Init(format.SampleRate, format.SampleRate.N(time.Second/10)); err != nil {
+		o.logger.Debug("failed to init speaker", "error", err)
+	}
+	done := make(chan bool)
+	o.mu.Lock()
+	o.currentDone = done
+	o.currentStream = &beep.Ctrl{Streamer: beep.Seq(streamer, beep.Callback(func() {
+		o.mu.Lock()
+		close(done)
+		o.currentStream = nil
+		o.currentDone = nil
+		o.mu.Unlock()
+	})), Paused: false}
+	o.mu.Unlock()
+	speaker.Play(o.currentStream)
+	<-done
+	return nil
+}
+
+func (o *KokoroOrator) Stop() {
+	// speaker.Clear()
+	o.logger.Debug("attempted to stop orator", "orator", o)
+	speaker.Lock()
+	defer speaker.Unlock()
+	o.mu.Lock()
+	defer o.mu.Unlock()
+	if o.currentStream != nil {
+		// o.currentStream.Paused = true
+		o.currentStream.Streamer = nil
+	}
+}
+
+func (o *GoogleTranslateOrator) stoproutine() {
+	for {
+		<-TTSDoneChan
+		o.logger.Debug("orator got done signal")
+		o.Stop()
+		// drain the channel
+		for len(TTSTextChan) > 0 {
+			<-TTSTextChan
+		}
+		o.mu.Lock()
+		o.textBuffer.Reset()
+		if o.currentDone != nil {
+			select {
+			case o.currentDone <- true:
+			default:
+				// Channel might be closed, ignore
+			}
+		}
+		o.interrupt = true
+		o.mu.Unlock()
+	}
+}
+
+func (o *GoogleTranslateOrator) readroutine() {
+	tokenizer, _ := english.NewSentenceTokenizer(nil)
+	for {
+		select {
+		case chunk := <-TTSTextChan:
+			o.mu.Lock()
+			o.interrupt = false
+			_, err := o.textBuffer.WriteString(chunk)
+			if err != nil {
+				o.logger.Warn("failed to write to stringbuilder", "error", err)
+				o.mu.Unlock()
+				continue
+			}
+			text := o.textBuffer.String()
+			sentences := tokenizer.Tokenize(text)
+			o.logger.Debug("adding chunk", "chunk", chunk, "text", text, "sen-len", len(sentences))
+			if len(sentences) <= 1 {
+				o.mu.Unlock()
+				continue
+			}
+			completeSentences := sentences[:len(sentences)-1]
+			remaining := sentences[len(sentences)-1].Text
+			o.textBuffer.Reset()
+			o.textBuffer.WriteString(remaining)
+			o.mu.Unlock()
+
+			for _, sentence := range completeSentences {
+				o.mu.Lock()
+				interrupted := o.interrupt
+				o.mu.Unlock()
+				if interrupted {
+					return
+				}
+				cleanedText := models.CleanText(sentence.Text)
+				if cleanedText == "" {
+					continue
+				}
+				o.logger.Debug("calling Speak with sentence", "sent", cleanedText)
+				if err := o.Speak(cleanedText); err != nil {
+					o.logger.Error("tts failed", "sentence", cleanedText, "error", err)
+				}
+			}
+		case <-TTSFlushChan:
+			o.logger.Debug("got flushchan signal start")
+			// lln is done get the whole message out
+			if len(TTSTextChan) > 0 { // otherwise might get stuck
+				for chunk := range TTSTextChan {
+					o.mu.Lock()
+					_, err := o.textBuffer.WriteString(chunk)
+					o.mu.Unlock()
+					if err != nil {
+						o.logger.Warn("failed to write to stringbuilder", "error", err)
+						continue
+					}
+					if len(TTSTextChan) == 0 {
+						break
+					}
+				}
+			}
+			o.mu.Lock()
+			remaining := o.textBuffer.String()
+			remaining = models.CleanText(remaining)
+			o.textBuffer.Reset()
+			o.mu.Unlock()
+			if remaining == "" {
+				continue
+			}
+			o.logger.Debug("calling Speak with remainder", "rem", remaining)
+			sentencesRem := tokenizer.Tokenize(remaining)
+			for _, rs := range sentencesRem { // to avoid dumping large volume of text
+				o.mu.Lock()
+				interrupt := o.interrupt
+				o.mu.Unlock()
+				if interrupt {
+					break
+				}
+				if err := o.Speak(rs.Text); err != nil {
+					o.logger.Error("tts failed", "sentence", rs.Text, "error", err)
+				}
+			}
+		}
+	}
+}
+
+func (o *GoogleTranslateOrator) GetLogger() *slog.Logger {
+	return o.logger
+}
+
+func (o *GoogleTranslateOrator) Speak(text string) error {
+	o.logger.Debug("fn: Speak is called", "text-len", len(text))
+	// Generate MP3 data using google-translate-tts
+	reader, err := o.speech.GenerateSpeech(text)
+	if err != nil {
+		o.logger.Error("generate speech failed", "error", err)
+		return fmt.Errorf("generate speech failed: %w", err)
+	}
+	// Decode the mp3 audio from reader (wrap with NopCloser for io.ReadCloser)
+	streamer, format, err := mp3.Decode(io.NopCloser(reader))
+	if err != nil {
+		o.logger.Error("mp3 decode failed", "error", err)
+		return fmt.Errorf("mp3 decode failed: %w", err)
+	}
+	defer streamer.Close()
+	playbackStreamer := beep.Streamer(streamer)
+	speed := o.speech.Speed
+	if speed <= 0 {
+		speed = 1.0
+	}
+	if speed != 1.0 {
+		playbackStreamer = beep.ResampleRatio(3, float64(speed), streamer)
+	}
+	// Initialize speaker with the format's sample rate
+	if err := speaker.Init(format.SampleRate, format.SampleRate.N(time.Second/10)); err != nil {
+		o.logger.Debug("failed to init speaker", "error", err)
+	}
+	done := make(chan bool)
+	o.mu.Lock()
+	o.currentDone = done
+	o.currentStream = &beep.Ctrl{Streamer: beep.Seq(playbackStreamer, beep.Callback(func() {
+		o.mu.Lock()
+		close(done)
+		o.currentStream = nil
+		o.currentDone = nil
+		o.mu.Unlock()
+	})), Paused: false}
+	o.mu.Unlock()
+	speaker.Play(o.currentStream)
+	<-done // wait for playback to complete
+	return nil
+}
+
+func (o *GoogleTranslateOrator) Stop() {
+	o.logger.Debug("attempted to stop google translate orator")
+	speaker.Lock()
+	defer speaker.Unlock()
+	o.mu.Lock()
+	defer o.mu.Unlock()
+	if o.currentStream != nil {
+		o.currentStream.Streamer = nil
+	}
+	// Also stop the speech handler if possible
+	if o.speech != nil {
+		_ = o.speech.Stop()
+	}
+}

go.mod

@@ -4,11 +4,12 @@ go 1.25.1
 
 require (
 	github.com/BurntSushi/toml v1.5.0
-	github.com/GrailFinder/google-translate-tts v0.1.4
+	github.com/GrailFinder/google-translate-tts v0.1.3
 	github.com/GrailFinder/searchagent v0.2.0
 	github.com/PuerkitoBio/goquery v1.11.0
 	github.com/gdamore/tcell/v2 v2.13.2
 	github.com/glebarez/go-sqlite v1.22.0
+	github.com/gopxl/beep/v2 v2.1.1
 	github.com/gordonklaus/portaudio v0.0.0-20250206071425-98a94950218b
 	github.com/jmoiron/sqlx v1.4.0
 	github.com/ledongthuc/pdf v0.0.0-20250511090121-5959a4027728
@@ -24,17 +25,21 @@ require (
 	github.com/andybalholm/cascadia v1.3.3 // indirect
 	github.com/deckarep/golang-set/v2 v2.8.0 // indirect
 	github.com/dustin/go-humanize v1.0.1 // indirect
+	github.com/ebitengine/oto/v3 v3.4.0 // indirect
+	github.com/ebitengine/purego v0.9.1 // indirect
 	github.com/emirpasic/gods v1.18.1 // indirect
 	github.com/gdamore/encoding v1.0.1 // indirect
 	github.com/go-jose/go-jose/v3 v3.0.4 // indirect
 	github.com/go-stack/stack v1.8.1 // indirect
 	github.com/google/uuid v1.6.0 // indirect
 	github.com/hajimehoshi/go-mp3 v0.3.4 // indirect
+	github.com/hajimehoshi/oto/v2 v2.3.1 // indirect
 	github.com/lucasb-eyer/go-colorful v1.3.0 // indirect
 	github.com/mattn/go-isatty v0.0.20 // indirect
 	github.com/mitchellh/colorstring v0.0.0-20190213212951-d06e56a500db // indirect
 	github.com/ncruces/go-strftime v1.0.0 // indirect
 	github.com/patrickmn/go-cache v2.1.0+incompatible // indirect
+	github.com/pkg/errors v0.9.1 // indirect
 	github.com/remyoudompheng/bigfft v0.0.0-20230129092748-24d4a6f8daec // indirect
 	github.com/rivo/uniseg v0.4.7 // indirect
 	github.com/schollz/progressbar/v2 v2.15.0 // indirect

go.sum

@@ -2,8 +2,8 @@ filippo.io/edwards25519 v1.1.0 h1:FNf4tywRC1HmFuKW5xopWpigGjJKiJSV0Cqo0cJWDaA=
 filippo.io/edwards25519 v1.1.0/go.mod h1:BxyFTGdWcka3PhytdK4V28tE5sGfRvvvRV7EaN4VDT4=
 github.com/BurntSushi/toml v1.5.0 h1:W5quZX/G/csjUnuI8SUYlsHs9M38FC7znL0lIO+DvMg=
 github.com/BurntSushi/toml v1.5.0/go.mod h1:ukJfTF/6rtPPRCnwkur4qwRxa8vTRFBF0uk2lLoLwho=
-github.com/GrailFinder/google-translate-tts v0.1.4 h1:NJoPZUGfBrmouQMN19MUcNPNUx4tmf4a8OZRME4E4Mg=
-github.com/GrailFinder/google-translate-tts v0.1.4/go.mod h1:YIOLKR7sObazdUCrSex3u9OVBovU55eYgWa25vsQJ18=
+github.com/GrailFinder/google-translate-tts v0.1.3 h1:Mww9tNzTWjjSh+OCbTPl/+21oMPKcUecXZfU7nTB/lA=
+github.com/GrailFinder/google-translate-tts v0.1.3/go.mod h1:YIOLKR7sObazdUCrSex3u9OVBovU55eYgWa25vsQJ18=
 github.com/GrailFinder/searchagent v0.2.0 h1:U2GVjLh/9xZt0xX9OcYk9Q2fMkyzyTiADPUmUisRdtQ=
 github.com/GrailFinder/searchagent v0.2.0/go.mod h1:d66tn5+22LI8IGJREUsRBT60P0sFdgQgvQRqyvgItrs=
 github.com/PuerkitoBio/goquery v1.11.0 h1:jZ7pwMQXIITcUXNH83LLk+txlaEy6NVOfTuP43xxfqw=
@@ -17,6 +17,10 @@ github.com/deckarep/golang-set/v2 v2.8.0 h1:swm0rlPCmdWn9mESxKOjWk8hXSqoxOp+Zlfu
 github.com/deckarep/golang-set/v2 v2.8.0/go.mod h1:VAky9rY/yGXJOLEDv3OMci+7wtDpOF4IN+y82NBOac4=
 github.com/dustin/go-humanize v1.0.1 h1:GzkhY7T5VNhEkwH0PVJgjz+fX1rhBrR7pRT3mDkpeCY=
 github.com/dustin/go-humanize v1.0.1/go.mod h1:Mu1zIs6XwVuF/gI1OepvI0qD18qycQx+mFykh5fBlto=
+github.com/ebitengine/oto/v3 v3.4.0 h1:br0PgASsEWaoWn38b2Goe7m1GKFYfNgnsjSd5Gg+/bQ=
+github.com/ebitengine/oto/v3 v3.4.0/go.mod h1:IOleLVD0m+CMak3mRVwsYY8vTctQgOM0iiL6S7Ar7eI=
+github.com/ebitengine/purego v0.9.1 h1:a/k2f2HQU3Pi399RPW1MOaZyhKJL9w/xFpKAg4q1s0A=
+github.com/ebitengine/purego v0.9.1/go.mod h1:iIjxzd6CiRiOG0UyXP+V1+jWqUXVjPKLAI0mRfJZTmQ=
 github.com/emirpasic/gods v1.18.1 h1:FXtiHYKDGKCW2KzwZKx0iC0PQmdlorYgdFG9jPXJ1Bc=
 github.com/emirpasic/gods v1.18.1/go.mod h1:8tpGGwCnJ5H4r6BWwaV6OrWmMoPhUl5jm/FMNAnJvWQ=
 github.com/gdamore/encoding v1.0.1 h1:YzKZckdBL6jVt2Gc+5p82qhrGiqMdG/eNs6Wy0u3Uhw=
@@ -37,10 +41,13 @@ github.com/google/pprof v0.0.0-20250317173921-a4b03ec1a45e h1:ijClszYn+mADRFY17k
 github.com/google/pprof v0.0.0-20250317173921-a4b03ec1a45e/go.mod h1:boTsfXsheKC2y+lKOCMpSfarhxDeIzfZG1jqGcPl3cA=
 github.com/google/uuid v1.6.0 h1:NIvaJDMOsjHA8n1jAhLSgzrAzy1Hgr+hNrb57e+94F0=
 github.com/google/uuid v1.6.0/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+yHo=
+github.com/gopxl/beep/v2 v2.1.1 h1:6FYIYMm2qPAdWkjX+7xwKrViS1x0Po5kDMdRkq8NVbU=
+github.com/gopxl/beep/v2 v2.1.1/go.mod h1:ZAm9TGQ9lvpoiFLd4zf5B1IuyxZhgRACMId1XJbaW0E=
 github.com/gordonklaus/portaudio v0.0.0-20250206071425-98a94950218b h1:WEuQWBxelOGHA6z9lABqaMLMrfwVyMdN3UgRLT+YUPo=
 github.com/gordonklaus/portaudio v0.0.0-20250206071425-98a94950218b/go.mod h1:esZFQEUwqC+l76f2R8bIWSwXMaPbp79PppwZ1eJhFco=
 github.com/hajimehoshi/go-mp3 v0.3.4 h1:NUP7pBYH8OguP4diaTZ9wJbUbk3tC0KlfzsEpWmYj68=
 github.com/hajimehoshi/go-mp3 v0.3.4/go.mod h1:fRtZraRFcWb0pu7ok0LqyFhCUrPeMsGRSVop0eemFmo=
+github.com/hajimehoshi/oto/v2 v2.3.1 h1:qrLKpNus2UfD674oxckKjNJmesp9hMh7u7QCrStB3Rc=
 github.com/hajimehoshi/oto/v2 v2.3.1/go.mod h1:seWLbgHH7AyUMYKfKYT9pg7PhUu9/SisyJvNTT+ASQo=
 github.com/hashicorp/golang-lru/v2 v2.0.7 h1:a+bsQ5rvGLjzHuww6tVxozPZFVghXaHOwFs4luLUK2k=
 github.com/hashicorp/golang-lru/v2 v2.0.7/go.mod h1:QeFd9opnmA6QUJc5vARoKUSoFhyfM2/ZepoAG6RGpeM=
@@ -64,6 +71,8 @@ github.com/neurosnap/sentences v1.1.2 h1:iphYOzx/XckXeBiLIUBkPu2EKMJ+6jDbz/sLJZ7
 github.com/neurosnap/sentences v1.1.2/go.mod h1:/pwU4E9XNL21ygMIkOIllv/SMy2ujHwpf8GQPu1YPbQ=
 github.com/patrickmn/go-cache v2.1.0+incompatible h1:HRMgzkcYKYpi3C8ajMPV8OFXaaRUnok+kx1WdO15EQc=
 github.com/patrickmn/go-cache v2.1.0+incompatible/go.mod h1:3Qf8kWWT7OJRJbdiICTKqZju1ZixQ/KpMGzzAfe6+WQ=
+github.com/pkg/errors v0.9.1 h1:FEBLx1zS214owpjy7qsBeixbURkuhQAwrK5UwLGTwt4=
+github.com/pkg/errors v0.9.1/go.mod h1:bwawxfHBFNV+L2hUp1rHADufV3IMtnDRdf1r5NINEl0=
 github.com/playwright-community/playwright-go v0.5700.1 h1:PNFb1byWqrTT720rEO0JL88C6Ju0EmUnR5deFLvtP/U=
 github.com/playwright-community/playwright-go v0.5700.1/go.mod h1:MlSn1dZrx8rszbCxY6x3qK89ZesJUYVx21B2JnkoNF0=
 github.com/pmezard/go-difflib v1.0.0 h1:4DBwDE0NGyQoBHbLQYPwSUPoCMWR5BEzIk/f1lZbAQM=

rag/embedder.go

@@ -7,6 +7,7 @@ import (
 	"fmt"
 	"gf-lt/config"
 	"gf-lt/models"
+	"gf-lt/onnx"
 	"log/slog"
 	"net/http"
 	"os"
@@ -156,43 +157,6 @@ type ONNXEmbedder struct {
 	modelPath     string
 }
 
-var onnxInitOnce sync.Once
-var onnxReady bool
-var onnxLibPath string
-var cudaLibPath string
-
-var onnxLibPaths = []string{
-	"/usr/lib/libonnxruntime.so",
-	"/usr/lib/libonnxruntime.so.1.24.2",
-	"/usr/local/lib/libonnxruntime.so",
-	"/usr/lib/x86_64-linux-gnu/libonnxruntime.so",
-	"/opt/onnxruntime/lib/libonnxruntime.so",
-}
-
-var cudaLibPaths = []string{
-	"/usr/lib/libonnxruntime_providers_cuda.so",
-	"/usr/local/lib/libonnxruntime_providers_cuda.so",
-	"/opt/onnxruntime/lib/libonnxruntime_providers_cuda.so",
-}
-
-func findONNXLibrary() string {
-	for _, path := range onnxLibPaths {
-		if _, err := os.Stat(path); err == nil {
-			return path
-		}
-	}
-	return ""
-}
-
-func findCUDALibrary() string {
-	for _, path := range cudaLibPaths {
-		if _, err := os.Stat(path); err == nil {
-			return path
-		}
-	}
-	return ""
-}
-
 func NewONNXEmbedder(modelPath, tokenizerPath string, dims int, logger *slog.Logger) (*ONNXEmbedder, error) {
 	// Check if model and tokenizer files exist
 	if _, err := os.Stat(modelPath); err != nil {
@@ -202,17 +166,16 @@ func NewONNXEmbedder(modelPath, tokenizerPath string, dims int, logger *slog.Log
 		return nil, fmt.Errorf("tokenizer not found: %w", err)
 	}
 
-	// Find ONNX library
-	onnxLibPath = findONNXLibrary()
-	if onnxLibPath == "" {
-		return nil, errors.New("ONNX runtime library not found in standard locations")
+	// Initialize ONNX runtime
+	if err := onnx.Init(); err != nil {
+		return nil, fmt.Errorf("ONNX init failed: %w", err)
+	}
+	if onnx.HasCUDASupport() {
+		logger.Info("ONNX CUDA support enabled")
+	} else {
+		logger.Info("ONNX using CPU fallback")
 	}
 
-	// Find CUDA provider library (optional)
-	cudaLibPath = findCUDALibrary()
-	if cudaLibPath == "" {
-		fmt.Println("WARNING: CUDA provider library not found, will use CPU")
-	}
 	emb := &ONNXEmbedder{
 		tokenizerPath: tokenizerPath,
 		dims:          dims,
@@ -239,26 +202,12 @@ func (e *ONNXEmbedder) ensureInitialized() error {
 		}
 		e.tokenizer = tok
 	}
-	onnxInitOnce.Do(func() {
-		onnxruntime_go.SetSharedLibraryPath(onnxLibPath)
-		if err := onnxruntime_go.InitializeEnvironment(); err != nil {
-			e.logger.Error("failed to initialize ONNX runtime", "error", err)
-			onnxReady = false
-			return
-		}
-		// Register CUDA provider if available
-		if cudaLibPath != "" {
-			if err := onnxruntime_go.RegisterExecutionProviderLibrary("CUDA", cudaLibPath); err != nil {
-				e.logger.Warn("failed to register CUDA provider", "error", err)
-			}
-		}
-		onnxReady = true
-	})
-	if !onnxReady {
+	// ONNX runtime already initialized by onnx.Init() in NewONNXEmbedder
+	if !onnx.IsReady() {
 		return errors.New("ONNX runtime not ready")
 	}
 	// Create session options
-	opts, err := onnxruntime_go.NewSessionOptions()
+	opts, err := onnx.NewSessionOptions()
 	if err != nil {
 		return fmt.Errorf("failed to create session options: %w", err)
 	}
@@ -266,27 +215,7 @@ func (e *ONNXEmbedder) ensureInitialized() error {
 		_ = opts.Destroy()
 	}()
 
-	// Try to add CUDA provider
-	useCUDA := cudaLibPath != ""
-	if useCUDA {
-		cudaOpts, err := onnxruntime_go.NewCUDAProviderOptions()
-		if err != nil {
-			e.logger.Warn("failed to create CUDA provider options, falling back to CPU", "error", err)
-			useCUDA = false
-		} else {
-			defer func() {
-				_ = cudaOpts.Destroy()
-			}()
-			if err := cudaOpts.Update(map[string]string{"device_id": "0"}); err != nil {
-				e.logger.Warn("failed to update CUDA options, falling back to CPU", "error", err)
-				useCUDA = false
-			} else if err := opts.AppendExecutionProviderCUDA(cudaOpts); err != nil {
-				e.logger.Warn("failed to append CUDA provider, falling back to CPU", "error", err)
-				useCUDA = false
-			}
-		}
-	}
-	if useCUDA {
+	if onnx.HasCUDASupport() {
 		e.logger.Info("Using CUDA for ONNX inference")
 	} else {
 		e.logger.Info("Using CPU for ONNX inference")