Add built-in phoneme frontend metadata

README.md

@@ -22,6 +22,10 @@ synthesize_cpp:
   quality_evaluation: not_run
   sample_rate_hz: 22050
   voice_count: 1
+  input_kinds:
+    - phonemes_utf8
+    - token_ids
+  frontend_provider: synthesize.symbol_map
   profiles:
     - F32
     - F16
@@ -48,9 +52,9 @@ Voice without requiring Voice selection.
 
 | Profile | Download | Size | Tensor storage | SHA-256 |
 | --- | --- | ---: | --- | --- |
-| F32 | [vits-ljspeech-F32.gguf](https://huggingface.co/jiangzhuo9357/vits-ljspeech-gguf/resolve/main/vits-ljspeech-F32.gguf) | 113.2 MB (113,244,864 bytes) | 460 F32 | `44d5fbfab510ad39b26a5224d77cf9e7d7cd8e2e520bdbdf76a42d62e4f929aa` |
-| F16 | [vits-ljspeech-F16.gguf](https://huggingface.co/jiangzhuo9357/vits-ljspeech-gguf/resolve/main/vits-ljspeech-F16.gguf) | 70.5 MB (70,453,376 bytes) | 342 F32 + 118 F16 | `ac14fd237532c75e485649e965009ddbfedc20ddbc69c157329abc11a2940d31` |
-| Q8_MIXED | [vits-ljspeech-Q8_MIXED.gguf](https://huggingface.co/jiangzhuo9357/vits-ljspeech-gguf/resolve/main/vits-ljspeech-Q8_MIXED.gguf) | 52.9 MB (52,890,048 bytes) | 342 F32 + 4 F16 + 114 Q8_0 | `ac026b3731ac627c44043bfd542933a0d277c6a1ac72f9e473b9334208308302` |
+| F32 | [vits-ljspeech-F32.gguf](https://huggingface.co/jiangzhuo9357/vits-ljspeech-gguf/resolve/main/vits-ljspeech-F32.gguf) | 113.2 MB (113,245,056 bytes) | 460 F32 | `bd17e44c7c2d761d33c1527059bd3921f9d3fa9d46bd73b3e020d746a8b7db7b` |
+| F16 | [vits-ljspeech-F16.gguf](https://huggingface.co/jiangzhuo9357/vits-ljspeech-gguf/resolve/main/vits-ljspeech-F16.gguf) | 70.5 MB (70,453,568 bytes) | 342 F32 + 118 F16 | `5fc428ba97416cc164055f509af1b9e120b089bd0bab792ad674c862411d7bca` |
+| Q8_MIXED | [vits-ljspeech-Q8_MIXED.gguf](https://huggingface.co/jiangzhuo9357/vits-ljspeech-gguf/resolve/main/vits-ljspeech-Q8_MIXED.gguf) | 52.9 MB (52,890,240 bytes) | 342 F32 + 4 F16 + 114 Q8_0 | `df95091f975e78088c4908c3f2adfc381cfa234c3f520ddd3ffe10160ff12ba1` |
 
 All profiles use the same VITS architecture and public synthesize.cpp API. The
 profile name describes a versioned storage policy, not the language or Execution
@@ -66,7 +70,7 @@ case. CUDA placement contained zero executable CPU fallback nodes.
 
 | Profile | CPU max PCM drift | DGX Spark CUDA max PCM drift | RTX 4070 SUPER CUDA max PCM drift |
 | --- | ---: | ---: | ---: |
-| F32 | 0.0002223924 | 0.0007436508 | 0.0009770896 |
+| F32 | 0.0002223924 | 0.20982037 | 0.2027478628 |
 | F16 | 0.0171255 | 0.20995625 | 0.20820463 |
 | Q8_MIXED | 0.36679696 | 0.329434 | 0.34679114 |
 
@@ -82,9 +86,15 @@ This package has one fixed package-default Voice and does not expose a preset
 speaker catalog. It produces 22050 Hz mono F32
 audio. Callers omit Voice selection; the package always uses its trained Voice.
 
-The current VITS package consumes token IDs produced by the pinned upstream
-`english_cleaners2` frontend. The portable runtime does not silently run eSpeak
-or download a frontend.
+This package accepts UTF-8 phoneme strings through the built-in
+`synthesize.symbol_map` frontend, and also accepts exact token IDs. The frontend
+validates UTF-8, maps each Unicode scalar through the symbol table embedded in
+the GGUF, and applies the model's blank-insertion rule.
+
+The built-in frontend does not perform grapheme-to-phoneme conversion or text
+normalization. Callers starting from raw text must currently run a compatible
+G2P frontend externally. The runtime does not silently invoke eSpeak or download
+a frontend.
 
 ## Usage
 
@@ -102,7 +112,7 @@ hf download jiangzhuo9357/vits-ljspeech-gguf vits-ljspeech-F16.gguf \
 build/bin/synthesize-cli \
   --model models/vits-ljspeech/vits-ljspeech-F16.gguf \
   --output output.wav \
-  --token-ids "0,156,0,47,0,102,0,4,0" \
+  --phonemes "ˈeɪ." \
   --language en \
   --seed 0
 ```

vits-ljspeech-F16.gguf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ac14fd237532c75e485649e965009ddbfedc20ddbc69c157329abc11a2940d31
-size 70453376
+oid sha256:5fc428ba97416cc164055f509af1b9e120b089bd0bab792ad674c862411d7bca
+size 70453568

vits-ljspeech-F32.gguf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:44d5fbfab510ad39b26a5224d77cf9e7d7cd8e2e520bdbdf76a42d62e4f929aa
-size 113244864
+oid sha256:bd17e44c7c2d761d33c1527059bd3921f9d3fa9d46bd73b3e020d746a8b7db7b
+size 113245056

vits-ljspeech-Q8_MIXED.gguf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ac026b3731ac627c44043bfd542933a0d277c6a1ac72f9e473b9334208308302
-size 52890048
+oid sha256:df95091f975e78088c4908c3f2adfc381cfa234c3f520ddd3ffe10160ff12ba1
+size 52890240