fix models

app.py

@@ -33,9 +33,11 @@ import torchaudio
 
 from examples import examples
 from model import (
+    COHERE_TRANSCRIBE_LANGUAGE_CODES,
     decode,
     get_pretrained_model,
     get_punct_model,
+    is_cohere_transcribe_model,
     language_to_models,
     sample_rate,
 )
@@ -77,6 +79,7 @@ def build_html_output(s: str, style: str = "result_item_success"):
 def process_url(
     language: str,
     repo_id: str,
+    cohere_language: str,
     decoding_method: str,
     num_active_paths: int,
     add_punct: str,
@@ -91,6 +94,7 @@ def process_url(
                 in_filename=f.name,
                 language=language,
                 repo_id=repo_id,
+                cohere_language=cohere_language,
                 decoding_method=decoding_method,
                 num_active_paths=num_active_paths,
                 add_punct=add_punct,
@@ -103,6 +107,7 @@ def process_url(
 def process_uploaded_file(
     language: str,
     repo_id: str,
+    cohere_language: str,
     decoding_method: str,
     num_active_paths: int,
     add_punct: str,
@@ -121,6 +126,7 @@ def process_uploaded_file(
             in_filename=in_filename,
             language=language,
             repo_id=repo_id,
+            cohere_language=cohere_language,
             decoding_method=decoding_method,
             num_active_paths=num_active_paths,
             add_punct=add_punct,
@@ -133,6 +139,7 @@ def process_uploaded_file(
 def process_microphone(
     language: str,
     repo_id: str,
+    cohere_language: str,
     decoding_method: str,
     num_active_paths: int,
     add_punct: str,
@@ -152,6 +159,7 @@ def process_microphone(
             in_filename=in_filename,
             language=language,
             repo_id=repo_id,
+            cohere_language=cohere_language,
             decoding_method=decoding_method,
             num_active_paths=num_active_paths,
             add_punct=add_punct,
@@ -165,6 +173,7 @@ def process_microphone(
 def process(
     language: str,
     repo_id: str,
+    cohere_language: str,
     decoding_method: str,
     num_active_paths: int,
     add_punct: str,
@@ -172,6 +181,7 @@ def process(
 ):
     MyPrint(f"language: {language}")
     MyPrint(f"repo_id: {repo_id}")
+    MyPrint(f"cohere_language: {cohere_language}")
     MyPrint(f"decoding_method: {decoding_method}")
     MyPrint(f"num_active_paths: {num_active_paths}")
     MyPrint(f"in_filename: {in_filename}")
@@ -204,7 +214,8 @@ def process(
         num_active_paths=num_active_paths,
     )
 
-    text = decode(recognizer, filename)
+    stream_language = cohere_language if is_cohere_transcribe_model(repo_id) else None
+    text = decode(recognizer, filename, language=stream_language)
     if add_punct == "Yes" and language in ("Chinese", "四川话"):
         punct = get_punct_model()
         text = punct.add_punctuation(text)
@@ -291,6 +302,10 @@ def update_model_dropdown(language: str):
     raise ValueError(f"Unsupported language: {language}")
 
 
+def update_cohere_language_visibility(repo_id: str):
+    return gr.update(visible=is_cohere_transcribe_model(repo_id))
+
+
 demo = gr.Blocks(css=css)
 
 
@@ -308,12 +323,31 @@ with demo:
         label="Select a model",
         value=language_to_models[language_choices[0]][0],
     )
+    cohere_language_dropdown = gr.Dropdown(
+        choices=COHERE_TRANSCRIBE_LANGUAGE_CODES,
+        value="en",
+        label="Cohere Transcribe language code",
+        info="Shown only for Cohere Transcribe. Select the language of the input audio.",
+        visible=False,
+    )
 
     language_radio.change(
         update_model_dropdown,
         inputs=language_radio,
         outputs=model_dropdown,
     )
+    language_radio.change(
+        lambda language: gr.update(
+            visible=is_cohere_transcribe_model(language_to_models[language][0])
+        ),
+        inputs=language_radio,
+        outputs=cohere_language_dropdown,
+    )
+    model_dropdown.change(
+        update_cohere_language_visibility,
+        inputs=model_dropdown,
+        outputs=cohere_language_dropdown,
+    )
 
     decoding_method_radio = gr.Radio(
         label="Decoding method",
@@ -402,6 +436,7 @@ with demo:
             inputs=[
                 language_radio,
                 model_dropdown,
+                cohere_language_dropdown,
                 decoding_method_radio,
                 num_active_paths_slider,
                 punct_radio,
@@ -415,6 +450,7 @@ with demo:
             inputs=[
                 language_radio,
                 model_dropdown,
+                cohere_language_dropdown,
                 decoding_method_radio,
                 num_active_paths_slider,
                 punct_radio,
@@ -428,6 +464,7 @@ with demo:
             inputs=[
                 language_radio,
                 model_dropdown,
+                cohere_language_dropdown,
                 decoding_method_radio,
                 num_active_paths_slider,
                 punct_radio,

model.py

@@ -16,7 +16,7 @@
 
 import os
 from functools import lru_cache
-from typing import Union
+from typing import Optional, Union
 
 from pathlib import Path
 import torch
@@ -113,8 +113,11 @@ def decode_online_recognizer(
 def decode_offline_recognizer_sherpa_onnx(
     recognizer: sherpa_onnx.OfflineRecognizer,
     filename: str,
+    language: Optional[str] = None,
 ) -> str:
     s = recognizer.create_stream()
+    if language:
+        s.set_option("language", language)
     samples, sample_rate = read_wave(filename)
     s.accept_waveform(sample_rate, samples)
     recognizer.decode_stream(s)
@@ -154,19 +157,40 @@ def decode(
         sherpa_onnx.OnlineRecognizer,
     ],
     filename: str,
+    language: Optional[str] = None,
 ) -> str:
     if isinstance(recognizer, sherpa.OfflineRecognizer):
         return decode_offline_recognizer(recognizer, filename)
     elif isinstance(recognizer, sherpa.OnlineRecognizer):
         return decode_online_recognizer(recognizer, filename)
     elif isinstance(recognizer, sherpa_onnx.OfflineRecognizer):
-        return decode_offline_recognizer_sherpa_onnx(recognizer, filename)
+        return decode_offline_recognizer_sherpa_onnx(
+            recognizer, filename, language=language
+        )
     elif isinstance(recognizer, sherpa_onnx.OnlineRecognizer):
         return decode_online_recognizer_sherpa_onnx(recognizer, filename)
     else:
         raise ValueError(f"Unknown recognizer type {type(recognizer)}")
 
 
+COHERE_TRANSCRIBE_LANGUAGE_CODES = [
+    "en",
+    "fr",
+    "de",
+    "it",
+    "es",
+    "pt",
+    "el",
+    "nl",
+    "pl",
+    "zh",
+    "ja",
+    "ko",
+    "vi",
+    "ar",
+]
+
+
 @lru_cache(maxsize=30)
 def get_pretrained_model(
     repo_id: str,
@@ -2562,6 +2586,10 @@ cohere_transcribe_14_languages_models = {
     "csukuangfj2/sherpa-onnx-cohere-transcribe-14-lang-int8-2026-04-01": _get_cohere_transcribe,
 }
 
+
+def is_cohere_transcribe_model(repo_id: str) -> bool:
+    return repo_id in cohere_transcribe_14_languages_models
+
 qwen3_asr_52_languages_models = {
     "csukuangfj2/sherpa-onnx-qwen3-asr-0.6B-int8-2026-03-25": _get_qwen3_asr,
 }

requirements.txt

@@ -10,7 +10,7 @@ numpy<2
 
 huggingface_hub
 
-# https://huggingface.co/csukuangfj2/sherpa-onnx-wheels/resolve/main/cpu/1.12.27/sherpa_onnx_core-1.12.27-py3-none-manylinux2014_x86_64.whl
-# https://huggingface.co/csukuangfj2/sherpa-onnx-wheels/resolve/main/cpu/1.12.27/sherpa_onnx-1.12.27-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.whl
+https://huggingface.co/csukuangfj2/sherpa-onnx-wheels/resolve/main/cpu/1.12.35/sherpa_onnx_core-1.12.35-py3-none-manylinux2014_x86_64.whl
+https://huggingface.co/csukuangfj2/sherpa-onnx-wheels/resolve/main/cpu/1.12.35/sherpa_onnx-1.12.35-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.whl
 
-sherpa-onnx>=1.12.35
+# sherpa-onnx>=1.12.35