home-services/tts-gateway/internal/adapters/secondary/jtalk/jtalk.go

// Package jtalk implements driven.TextNormalizer by shelling out to the
// open_jtalk CLI, reproducing uma-tts-api's japanese_cleaners pipeline
// (text/cleaners.py) exactly - verified rune-for-rune against the real
// Python implementation during TTS_GATEWAY_PLAN.md's Phase 0 spike. See
// tmp/reference/uma-tts-api/spike/FINDINGS.md for the verification detail.
package jtalk

import (
	"bytes"
	"context"
	"fmt"
	"io/fs"
	"os"
	"os/exec"
	"path/filepath"
	"regexp"
	"strings"

	"gitea.nik4nao.com/nik/home-services/tts-gateway/internal/config"
	"gitea.nik4nao.com/nik/home-services/tts-gateway/internal/core/domain"
)

type Client struct {
	bin     string
	dictDir string
	voice   string
}

// NewClient constructs the open_jtalk CLI adapter, auto-discovering the
// dictionary directory and voice file when the config leaves them empty
// (matches how the Phase 0 spike located them on nik-gpu's open-jtalk +
// open-jtalk-mecab-naist-jdic + hts-voice-* apt packages).
func NewClient(cfg *config.Config) (*Client, error) {
	dictDir := cfg.OpenJTalkDictDir
	if dictDir == "" {
		var err error
		dictDir, err = findDictDir()
		if err != nil {
			return nil, err
		}
	}

	voice := cfg.OpenJTalkVoice
	if voice == "" {
		var err error
		voice, err = findVoice()
		if err != nil {
			return nil, err
		}
	}

	return &Client{bin: cfg.OpenJTalkBin, dictDir: dictDir, voice: voice}, nil
}

var dictDirCandidates = []string{
	"/usr/lib/*/open_jtalk/open_jtalk_dic_utf_8-*",
	"/usr/share/open_jtalk/dic/*",
	"/var/lib/mecab/dic/open-jtalk/naist-jdic",
}

func findDictDir() (string, error) {
	for _, pattern := range dictDirCandidates {
		matches, err := filepath.Glob(pattern)
		if err != nil {
			continue
		}
		for _, m := range matches {
			if info, err := os.Stat(m); err == nil && info.IsDir() {
				return m, nil
			}
		}
	}
	return "", fmt.Errorf("no open_jtalk dictionary dir found under %v", dictDirCandidates)
}

func findVoice() (string, error) {
	const root = "/usr/share/hts-voice"
	var found string
	err := filepath.WalkDir(root, func(path string, d fs.DirEntry, err error) error {
		if err != nil || found != "" {
			return nil
		}
		if !d.IsDir() && strings.HasSuffix(path, ".htsvoice") {
			found = path
		}
		return nil
	})
	if err != nil || found == "" {
		return "", fmt.Errorf("no .htsvoice file found under %s", root)
	}
	return found, nil
}

// Normalize reproduces japanese_cleaners(text) then maps the result
// character-by-character to symbol IDs (text_to_sequence's actual
// behavior - see FINDINGS.md on why this is intentional-if-buggy fidelity,
// not a mistake), and finally intersperses blank tokens (add_blank=true).
func (c *Client) Normalize(ctx context.Context, text string) ([]int64, error) {
	cleaned, err := c.japaneseCleaners(ctx, text)
	if err != nil {
		return nil, err
	}

	seq := make([]int64, 0, len(cleaned))
	for _, r := range cleaned {
		id, ok := domain.SymbolToID[r]
		if !ok {
			return nil, fmt.Errorf("jtalk: no symbol for rune %q in cleaned text %q", r, cleaned)
		}
		seq = append(seq, int64(id))
	}

	return domain.Intersperse(seq, 0), nil
}

// japaneseCleaners ports text/cleaners.py::japanese_cleaners rune-for-rune.
func (c *Client) japaneseCleaners(ctx context.Context, text string) (string, error) {
	spans, marks := splitByMarks(text)

	var b strings.Builder
	for i, mark := range marks {
		if spans[i] != "" {
			phonemes, err := c.g2p(ctx, spans[i])
			if err != nil {
				return "", err
			}
			b.WriteString(phonemes)
		}
		b.WriteString(markToASCII(mark))
	}
	if last := spans[len(spans)-1]; last != "" {
		phonemes, err := c.g2p(ctx, last)
		if err != nil {
			return "", err
		}
		b.WriteString(phonemes)
	}

	out := b.String()
	if out != "" {
		r := []rune(out)
		last := r[len(r)-1]
		if (last >= 'A' && last <= 'Z') || (last >= 'a' && last <= 'z') {
			out += "."
		}
	}
	return out, nil
}

// splitByMarks reproduces re.split(_japanese_marks, text) / re.findall(...):
// every individual non-Japanese-classified rune is its own split point, so
// consecutive marks yield an empty span between them. len(spans) is always
// len(marks)+1.
func splitByMarks(text string) (spans []string, marks []rune) {
	var current strings.Builder
	for _, r := range text {
		if isJapaneseChar(r) {
			current.WriteRune(r)
			continue
		}
		spans = append(spans, current.String())
		marks = append(marks, r)
		current.Reset()
	}
	spans = append(spans, current.String())
	return spans, marks
}

// isJapaneseChar mirrors pyopenjtalk's cleaner regex
// `[A-Za-z\d々぀-ヿ一-鿿１-９Ａ-Ｚａ-ｚｦ-ﾝ]` exactly, including its quirks (e.g.
// fullwidth "０" (U+FF10) is NOT included, only "１"-"９").
func isJapaneseChar(r rune) bool {
	switch {
	case r >= 'A' && r <= 'Z', r >= 'a' && r <= 'z', r >= '0' && r <= '9':
		return true
	case r == '々': // U+3005
		return true
	case r >= '぀' && r <= 'ヿ': // ぀-ヿ: Hiragana + Katakana
		return true
	case r >= '一' && r <= '鿿': // 一-鿿: CJK Unified Ideographs
		return true
	case r >= '１' && r <= '９': // １-９ (fullwidth 1-9, NOT 0)
		return true
	case r >= 'Ａ' && r <= 'Ｚ': // Ａ-Ｚ
		return true
	case r >= 'ａ' && r <= 'ｚ': // ａ-ｚ
		return true
	case r >= 'ｦ' && r <= 'ﾝ': // ｦ-ﾝ (halfwidth katakana)
		return true
	default:
		return false
	}
}

// markToASCII reproduces unidecode(mark).replace(' ', ”) for the small set
// of punctuation marks _japanese_marks actually matches in practice (verified
// against real unidecode output during the Phase 0 spike). ASCII runes are
// unidecode no-ops so pass through unchanged; anything else not in this
// table is dropped (empty string) rather than guessing a transliteration.
func markToASCII(r rune) string {
	if r < 128 {
		return string(r)
	}
	if s, ok := markTable[r]; ok {
		return s
	}
	return ""
}

var markTable = map[rune]string{
	'、': ",",
	'。': ".",
	'「': "[",
	'」': "]",
	'『': "{",
	'』': "}",
	'・': "*",
	'〜': "~",
	'！': "!",
	'？': "?",
	'，': ",",
	'．': ".",
	'　': " ", // fullwidth space
	'…': "...",
	'―': "--",
}

var labelPhonemeRe = regexp.MustCompile(`-([^+]+)\+`)

// g2p runs the open_jtalk CLI on one Japanese-only span and returns the
// concatenated phoneme string with 'pau' tokens and spaces stripped, matching
// pyopenjtalk.g2p(span, kana=False).replace('pau',”).replace(' ',”).
func (c *Client) g2p(ctx context.Context, span string) (string, error) {
	cmd := exec.CommandContext(ctx, c.bin, "-x", c.dictDir, "-m", c.voice, "-ot", "/dev/stdout", "-ow", "/dev/null")
	cmd.Stdin = strings.NewReader(span)
	var stdout bytes.Buffer
	cmd.Stdout = &stdout
	if err := cmd.Run(); err != nil {
		return "", fmt.Errorf("jtalk: open_jtalk run: %w", err)
	}

	phonemes, err := parseOutputLabelPhonemes(stdout.String())
	if err != nil {
		return "", err
	}

	var b strings.Builder
	for _, p := range phonemes {
		if p == "pau" {
			continue
		}
		b.WriteString(p)
	}
	return b.String(), nil
}

// parseOutputLabelPhonemes extracts the current-phoneme field from each HTS
// full-context label line within open_jtalk's verbose "-ot /dev/stdout"
// output, scoped to only the [Output label] section - the rest of the
// verbose dump ([Global parameter] etc.) can contain lines that spuriously
// match the same "-X+" pattern, which silently corrupted an earlier version
// of this parser (see FINDINGS.md). Boundary sil tokens are stripped.
func parseOutputLabelPhonemes(rawStdout string) ([]string, error) {
	const startMarker = "[Output label]"
	_, section, ok := strings.Cut(rawStdout, startMarker)
	if !ok {
		return nil, fmt.Errorf("jtalk: %q not found in open_jtalk output", startMarker)
	}
	if end := strings.IndexByte(section, '['); end != -1 {
		section = section[:end]
	}

	var phonemes []string
	for line := range strings.SplitSeq(strings.TrimSpace(section), "\n") {
		m := labelPhonemeRe.FindStringSubmatch(line)
		if m != nil {
			phonemes = append(phonemes, m[1])
		}
	}
	if len(phonemes) > 0 && phonemes[0] == "sil" {
		phonemes = phonemes[1:]
	}
	if len(phonemes) > 0 && phonemes[len(phonemes)-1] == "sil" {
		phonemes = phonemes[:len(phonemes)-1]
	}
	return phonemes, nil
}