add matcha

app.py

@@ -71,6 +71,13 @@ css = """
 """
 
 examples = [
+    [
+        "Chinese (Mandarin, 普通话)",
+        "csukuangfj/matcha-icefall-zh-baker|1",
+        "某某银行的副行长和一些行政领导表示，他们去过长江和长白山; 经济不断增长。2024年12月31号，拨打110或者18920240511。123456块钱。",
+        0,
+        1.0,
+    ],
     [
         "Chinese (Mandarin, 普通话)",
         "csukuangfj/vits-zh-hf-fanchen-wnj|1",

model.py

@@ -234,6 +234,83 @@ def _get_vits_zh_aishell3(repo_id: str, speed: float) -> sherpa_onnx.OfflineTts:
     return tts
 
 
+@lru_cache(maxsize=10)
+def _get_matcha_hf(repo_id: str, speed: float) -> sherpa_onnx.OfflineTts:
+    repo_id = repo_id.split("|")[0]
+    assert repo_id in ("csukuangfj/matcha-icefall-zh-baker",)
+
+    if repo_id == "csukuangfj/matcha-icefall-zh-baker":
+        acoustic_model = "model-steps-3.onnx"
+
+    if not Path("/tmp/dict").is_dir():
+        os.system(
+            "cd /tmp; curl -SL -O https://github.com/csukuangfj/cppjieba/releases/download/sherpa-onnx-2024-04-19/dict.tar.bz2; tar xvf dict.tar.bz2"
+        )
+    os.system("ls -lh /tmp/dict")
+
+    acoustic_model = get_file(
+        repo_id=repo_id,
+        filename=acoustic_model,
+        subfolder=".",
+    )
+
+    vocoder = get_file(
+        repo_id="csukuangfj/sherpa-onnx-hifigan",
+        filename="hifigan_v2.onnx",
+        subfolder=".",
+    )
+
+    lexicon = get_file(
+        repo_id=repo_id,
+        filename="lexicon.txt",
+        subfolder=".",
+    )
+
+    tokens = get_file(
+        repo_id=repo_id,
+        filename="tokens.txt",
+        subfolder=".",
+    )
+
+    rule_fars = ""
+
+    rule_fsts = ["phone.fst", "date.fst", "number.fst"]
+
+    rule_fsts = [
+        get_file(
+            repo_id=repo_id,
+            filename=f,
+            subfolder=".",
+        )
+        for f in rule_fsts
+    ]
+    rule_fsts = ",".join(rule_fsts)
+
+    vits_dict_dir = "/tmp/dict"
+
+    tts_config = sherpa_onnx.OfflineTtsConfig(
+        model=sherpa_onnx.OfflineTtsModelConfig(
+            vits=sherpa_onnx.OfflineTtsMatchaModelConfig(
+                acoustic_model=acoustic_model,
+                vocoder=vocoder,
+                lexicon=lexicon,
+                tokens=tokens,
+                dict_dir=dict_dir,
+                length_scale=1.0 / speed,
+            ),
+            provider="cpu",
+            debug=True,
+            num_threads=2,
+        ),
+        rule_fsts=rule_fsts,
+        rule_fars=rule_fars,
+        max_num_sentences=1,
+    )
+    tts = sherpa_onnx.OfflineTts(tts_config)
+
+    return tts
+
+
 @lru_cache(maxsize=10)
 def _get_vits_hf(repo_id: str, speed: float) -> sherpa_onnx.OfflineTts:
     repo_id = repo_id.split("|")[0]
@@ -444,6 +521,7 @@ persian_english_models = {
 }
 
 chinese_models = {
+    "csukuangfj/matcha-icefall-zh-baker|1": _get_matcha_hf,  # 1
     "csukuangfj/vits-zh-hf-fanchen-wnj|1": _get_vits_hf,  # 1
     "csukuangfj/vits-zh-hf-fanchen-C|187": _get_vits_hf,  # 187
     "csukuangfj/sherpa-onnx-vits-zh-ll|5": _get_vits_hf,  # 804

requirements.txt

@@ -1,4 +1,4 @@
-https://huggingface.co/csukuangfj/sherpa-onnx-wheels/resolve/main/cpu/1.10.29/sherpa_onnx-1.10.29-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
-#sherpa-onnx>=1.10.16
+#https://huggingface.co/csukuangfj/sherpa-onnx-wheels/resolve/main/cpu/1.10.29/sherpa_onnx-1.10.29-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
+sherpa-onnx>=1.10.37
 
 soundfile