1. Add two models

sherpa-onnx-streaming-zipformer-zh-int8-2025-06-30
sherpa-onnx-streaming-zipformer-zh-xlarge-int8-2025-06-30
2. Secure weight file selection behavoir especially for int8 only model

app/asr_worker.py

@@ -18,6 +18,30 @@ to_ZHCN = OpenCC('t2s')
 
 # Streaming Zipformer model registry: paths relative to repo root
 STREAMING_ZIPFORMER_MODELS = {
+    # csukuangfj/sherpa-onnx-streaming-zipformer-zh-int8-2025-06-30
+    "csukuangfj/sherpa-onnx-streaming-zipformer-zh-int8-2025-06-30": {
+        "tokens": "tokens.txt",
+        "encoder_fp32": "encoder.int8.onnx",
+        "encoder_int8": None,
+        "decoder_fp32": "decoder.onnx",
+        "decoder_int8": None,
+        "joiner_fp32": "joiner.int8.onnx",
+        "joiner_int8": None,
+        "modeling_unit":"cjkchar",
+        "bpe_model":   None,
+    },
+    # csukuangfj/sherpa-onnx-streaming-zipformer-zh-xlarge-int8-2025-06-30
+    "csukuangfj/sherpa-onnx-streaming-zipformer-zh-xlarge-int8-2025-06-30": {
+        "tokens": "tokens.txt",
+        "encoder_fp32": "encoder.int8.onnx",
+        "encoder_int8": None,
+        "decoder_fp32": "decoder.onnx",
+        "decoder_int8": None,
+        "joiner_fp32": "joiner.int8.onnx",
+        "joiner_int8": None,
+        "modeling_unit":"cjkchar+bpe",
+        "bpe_model":   "bpe.model",
+    },
     # bilingual zh-en with char+BPE
     "csukuangfj/k2fsa-zipformer-bilingual-zh-en-t": {
         "tokens": "data/lang_char_bpe/tokens.txt",
@@ -177,6 +201,38 @@ STREAMING_ZIPFORMER_MODELS = {
 def resample_audio(audio: np.ndarray, orig_sr: int, target_sr: int) -> np.ndarray:
     return scipy.signal.resample_poly(audio, target_sr, orig_sr)
 
+def choose_file(entry: dict, component: str, precision: str) -> str | None:
+    """
+    Pick the best file for the given component ('encoder', 'decoder', or 'joiner'),
+    but only if it really exists on disk.
+
+    1) Look up the two candidates.
+    2) Discard any whose path doesn’t exist.
+    3) If neither exists, print an error.
+    4) If exactly one remains, return it.
+    5) Otherwise, fall back to “int8 if requested & available, else fp32”.
+    """
+    e8  = entry.get(f'{component}_int8')
+    e32 = entry.get(f'{component}_fp32')
+
+    # Sanity: drop any that aren’t real files
+    if e8 and not os.path.exists(e8):
+        e8 = None
+    if e32 and not os.path.exists(e32):
+        e32 = None
+
+    # If neither exists, error out
+    if e8 is None and e32 is None:
+        print(f"Error: no available files for component '{component}' (neither int8 nor fp32 found).")
+        return None
+
+    # If exactly one exists → pick it
+    if (e8 is None) != (e32 is None):
+        return e8 or e32
+
+    # Otherwise, fall back to “int8 if requested & available, else fp32”
+    return e8 if precision == 'int8' and e8 else e32
+
 # Create an online recognizer for a given model and precision
 # model_id: full HF repo ID
 # precision: "int8" or "fp32"
@@ -194,9 +250,9 @@ def create_recognizer(
     entry = STREAMING_ZIPFORMER_MODELS[model_id]
 
     tokens_file = entry['tokens']
-    encoder_file = entry['encoder_int8'] if precision == 'int8' and entry['encoder_int8'] else entry['encoder_fp32']
-    decoder_file = entry['decoder_int8'] if precision == 'int8' and entry['decoder_int8'] else entry['decoder_fp32']
-    joiner_file = entry['joiner_int8'] if precision == 'int8' and entry['joiner_int8'] else entry['joiner_fp32']
+    encoder_file = choose_file(entry, 'encoder', precision)
+    decoder_file = choose_file(entry, 'decoder', precision)
+    joiner_file  = choose_file(entry, 'joiner',  precision)
 
     tokens_path = hf_hub_download(repo_id=model_id, filename=tokens_file, cache_dir=str(CACHE_DIR))
     encoder_path = hf_hub_download(repo_id=model_id, filename=encoder_file, cache_dir=str(CACHE_DIR))