Add C++ runtime and Python API for Google MedASR models

.github/scripts/test-python.sh

@@ -8,6 +8,17 @@ log() {
   echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
 }
 
+log "test Google MedASR"
+curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-medasr-ctc-en-int8-2025-12-25.tar.bz2
+tar xvf sherpa-onnx-medasr-ctc-en-int8-2025-12-25.tar.bz2
+rm sherpa-onnx-medasr-ctc-en-int8-2025-12-25.tar.bz2
+ls -lh sherpa-onnx-medasr-ctc-en-int8-2025-12-25
+
+ls -lh sherpa-onnx-medasr-ctc-en-int8-2025-12-25/test_wavs
+
+python3 ./python-api-examples/offline-medasr-ctc-decode-files.py
+rm -rf sherpa-onnx-medasr-ctc-en-int8-2025-12-25
+
 log "test omnilingual ASR"
 curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-omnilingual-asr-1600-languages-300M-ctc-int8-2025-11-12.tar.bz2
 tar xvf sherpa-onnx-omnilingual-asr-1600-languages-300M-ctc-int8-2025-11-12.tar.bz2

.github/workflows/export-medasr-ctc-to-onnx.yaml

@@ -3,7 +3,7 @@ name: export-medasr-ctc-to-onnx
 on:
   push:
     branches:
-      - export-medasr-onnx
+      - cpp-medasr-2
   workflow_dispatch:
 
 concurrency:
@@ -42,7 +42,9 @@ jobs:
         run: |
           cd scripts/medasr
 
-          curl -SL -O https://huggingface.co/csukuangfj/sherpa-onnx-medasr-ctc-en-int8-2025-12-25/resolve/main/test_wavs/0.wav
+          for i in $(seq 0 5); do
+            curl -SL -O https://huggingface.co/csukuangfj/sherpa-onnx-medasr-ctc-en-int8-2025-12-25/resolve/main/test_wavs/$i.wav
+          done
 
           curl -SL -O https://huggingface.co/csukuangfj/sherpa-onnx-medasr-ctc-en-int8-2025-12-25/resolve/main/test_wavs/transcript.txt
 
@@ -53,7 +55,9 @@ jobs:
         run: |
           cd scripts/medasr
 
-          python3 test_onnx.py --model ./model.onnx --tokens ./tokens.txt --wav ./0.wav
+          for i in $(seq 0 5); do
+            python3 test_onnx.py --model ./model.onnx --tokens ./tokens.txt --wav ./$i.wav
+          done
 
           cat transcript.txt
 
@@ -62,7 +66,9 @@ jobs:
         run: |
           cd scripts/medasr
 
-          python3 test_onnx.py --model ./model.int8.onnx --tokens ./tokens.txt --wav ./0.wav
+          for i in $(seq 0 5); do
+            python3 test_onnx.py --model ./model.int8.onnx --tokens ./tokens.txt --wav ./$i.wav
+          done
 
           cat transcript.txt
 

python-api-examples/offline-medasr-ctc-decode-files.py

@@ -0,0 +1,142 @@
+#!/usr/bin/env python3
+
+"""
+This file shows how to use a non-streaming Google MedASR CTC model from
+https://huggingface.co/google/medasr
+to decode files.
+
+Please download model files from
+https://github.com/k2-fsa/sherpa-onnx/releases/tag/asr-models
+
+For instance,
+
+wget https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-medasr-ctc-en-int8-2025-12-25.tar.bz2
+tar xvf sherpa-onnx-medasr-ctc-en-int8-2025-12-25.tar.bz2
+rm sherpa-onnx-medasr-ctc-en-int8-2025-12-25.tar.bz2
+"""
+
+import time
+from pathlib import Path
+
+import librosa
+import numpy as np
+import sherpa_onnx
+
+
+def create_recognizer():
+    model = "./sherpa-onnx-medasr-ctc-en-int8-2025-12-25/model.int8.onnx"
+    tokens = "./sherpa-onnx-medasr-ctc-en-int8-2025-12-25/tokens.txt"
+    test_wav_0 = "./sherpa-onnx-medasr-ctc-en-int8-2025-12-25/test_wavs/0.wav"
+    test_wav_1 = "./sherpa-onnx-medasr-ctc-en-int8-2025-12-25/test_wavs/1.wav"
+    test_wav_2 = "./sherpa-onnx-medasr-ctc-en-int8-2025-12-25/test_wavs/2.wav"
+    test_wav_3 = "./sherpa-onnx-medasr-ctc-en-int8-2025-12-25/test_wavs/3.wav"
+    test_wav_4 = "./sherpa-onnx-medasr-ctc-en-int8-2025-12-25/test_wavs/4.wav"
+    test_wav_5 = "./sherpa-onnx-medasr-ctc-en-int8-2025-12-25/test_wavs/5.wav"
+
+    for f in [
+        model,
+        tokens,
+        test_wav_0,
+        test_wav_1,
+        test_wav_2,
+        test_wav_3,
+        test_wav_4,
+        test_wav_5,
+    ]:
+        if not Path(f).is_file():
+            print(f"{f} does not exist")
+
+            raise ValueError(
+                """Please download model files from
+                https://github.com/k2-fsa/sherpa-onnx/releases/tag/asr-models
+                """
+            )
+    return (
+        sherpa_onnx.OfflineRecognizer.from_medasr_ctc(
+            model=model,
+            tokens=tokens,
+            num_threads=2,
+        ),
+        test_wav_0,
+        test_wav_1,
+        test_wav_2,
+        test_wav_3,
+        test_wav_4,
+        test_wav_5,
+    )
+
+
+def load_audio(filename):
+    audio, sample_rate = librosa.load(filename, sr=16000)
+    assert sample_rate == 16000, sample_rate
+
+    return np.ascontiguousarray(audio)
+
+
+def decode_single_file(recognizer, filename):
+    samples = load_audio(filename)
+
+    start_time = time.time()
+
+    stream = recognizer.create_stream()
+    stream.accept_waveform(sample_rate=16000, waveform=samples)
+    recognizer.decode_stream(stream)
+
+    end_time = time.time()
+    elapsed_seconds = end_time - start_time
+    audio_duration = len(samples) / 16000
+    real_time_factor = elapsed_seconds / audio_duration
+
+    print("---")
+    print(filename)
+    print(stream.result)
+    print(f"Elapsed seconds: {elapsed_seconds:.3f}")
+    print(f"Audio duration in seconds: {audio_duration:.3f}")
+    print(f"RTF: {elapsed_seconds:.3f}/{audio_duration:.3f} = {real_time_factor:.3f}")
+    print()
+
+
+def decode_multiple_files(recognizer, filenames):
+    streams = []
+
+    start_time = time.time()
+
+    audio_duration = 0
+
+    for filename in filenames:
+        samples = load_audio(filename)
+        audio_duration += len(samples) / 16000
+
+        stream = recognizer.create_stream()
+        stream.accept_waveform(sample_rate=16000, waveform=samples)
+        streams.append(stream)
+
+    recognizer.decode_streams(streams)
+
+    end_time = time.time()
+    elapsed_seconds = end_time - start_time
+    real_time_factor = elapsed_seconds / audio_duration
+
+    for name, stream in zip(filenames, streams):
+        print("---")
+        print(name)
+        print(stream.result)
+        print()
+
+    print(f"Elapsed seconds: {elapsed_seconds:.3f}")
+    print(f"Audio duration in seconds: {audio_duration:.3f}")
+    print(f"RTF: {elapsed_seconds:.3f}/{audio_duration:.3f} = {real_time_factor:.3f}")
+    print()
+    print()
+
+
+def main():
+    recognizer, *filenames = create_recognizer()
+
+    decode_single_file(recognizer, filenames[0])
+    decode_single_file(recognizer, filenames[1])
+    decode_multiple_files(recognizer, filenames[2:])
+
+
+if __name__ == "__main__":
+    main()

scripts/medasr/export_onnx.py

@@ -107,6 +107,7 @@ def main():
         "model_author": "google",
         "maintainer": "k2-fsa",
         "vocab_size": processor.tokenizer.vocab_size,
+        "subsampling_factor": 4,
         "url": "https://github.com/Google-Health/medasr",
         "license": "https://developers.google.com/health-ai-developer-foundations/terms",
     }

sherpa-onnx/csrc/CMakeLists.txt

@@ -40,6 +40,8 @@ set(sources
   offline-fire-red-asr-model.cc
   offline-lm-config.cc
   offline-lm.cc
+  offline-medasr-ctc-model-config.cc
+  offline-medasr-ctc-model.cc
   offline-model-config.cc
   offline-moonshine-greedy-search-decoder.cc
   offline-moonshine-model-config.cc

sherpa-onnx/csrc/offline-ctc-model.cc

@@ -21,6 +21,7 @@
 #include "sherpa-onnx/csrc/file-utils.h"
 #include "sherpa-onnx/csrc/macros.h"
 #include "sherpa-onnx/csrc/offline-dolphin-model.h"
+#include "sherpa-onnx/csrc/offline-medasr-ctc-model.h"
 #include "sherpa-onnx/csrc/offline-nemo-enc-dec-ctc-model.h"
 #include "sherpa-onnx/csrc/offline-omnilingual-asr-ctc-model.h"
 #include "sherpa-onnx/csrc/offline-tdnn-ctc-model.h"
@@ -126,6 +127,8 @@ std::unique_ptr<OfflineCtcModel> OfflineCtcModel::Create(
     return std::make_unique<OfflineTeleSpeechCtcModel>(config);
   } else if (!config.omnilingual.model.empty()) {
     return std::make_unique<OfflineOmnilingualAsrCtcModel>(config);
+  } else if (!config.medasr.model.empty()) {
+    return std::make_unique<OfflineMedAsrCtcModel>(config);
   }
 
   // TODO(fangjun): Refactor it. We don't need to use model_type here
@@ -192,6 +195,8 @@ std::unique_ptr<OfflineCtcModel> OfflineCtcModel::Create(
     return std::make_unique<OfflineTeleSpeechCtcModel>(mgr, config);
   } else if (!config.omnilingual.model.empty()) {
     return std::make_unique<OfflineOmnilingualAsrCtcModel>(mgr, config);
+  } else if (!config.medasr.model.empty()) {
+    return std::make_unique<OfflineMedAsrCtcModel>(mgr, config);
   }
 
   // TODO(fangjun): Refactor it. We don't need to use model_type here

sherpa-onnx/csrc/offline-medasr-ctc-model-config.cc

@@ -0,0 +1,40 @@
+// sherpa-onnx/csrc/offline-medasr-ctc-model-config.cc
+//
+// Copyright (c)  2025  Xiaomi Corporation
+
+#include "sherpa-onnx/csrc/offline-medasr-ctc-model-config.h"
+
+#include <sstream>
+#include <string>
+
+#include "sherpa-onnx/csrc/file-utils.h"
+#include "sherpa-onnx/csrc/macros.h"
+
+namespace sherpa_onnx {
+
+void OfflineMedAsrCtcModelConfig::Register(ParseOptions *po) {
+  po->Register(
+      "medasr", &model,
+      "Path to model.onnx from MedASR. Please see "
+      "https://github.com/k2-fsa/sherpa-onnx/pull/2934 for available models");
+}
+
+bool OfflineMedAsrCtcModelConfig::Validate() const {
+  if (!FileExists(model)) {
+    SHERPA_ONNX_LOGE("MedASR model: '%s' does not exist", model.c_str());
+    return false;
+  }
+
+  return true;
+}
+
+std::string OfflineMedAsrCtcModelConfig::ToString() const {
+  std::ostringstream os;
+
+  os << "OfflineMedAsrCtcModelConfig(";
+  os << "model=\"" << model << "\")";
+
+  return os.str();
+}
+
+}  // namespace sherpa_onnx

sherpa-onnx/csrc/offline-medasr-ctc-model-config.h

@@ -0,0 +1,29 @@
+// sherpa-onnx/csrc/offline-medasr-ctc-model-config.h
+//
+// Copyright (c)  2025  Xiaomi Corporation
+
+#ifndef SHERPA_ONNX_CSRC_OFFLINE_MEDASR_CTC_MODEL_CONFIG_H_
+#define SHERPA_ONNX_CSRC_OFFLINE_MEDASR_CTC_MODEL_CONFIG_H_
+
+#include <string>
+
+#include "sherpa-onnx/csrc/parse-options.h"
+
+namespace sherpa_onnx {
+
+struct OfflineMedAsrCtcModelConfig {
+  std::string model;
+
+  OfflineMedAsrCtcModelConfig() = default;
+  explicit OfflineMedAsrCtcModelConfig(const std::string &model)
+      : model(model) {}
+
+  void Register(ParseOptions *po);
+  bool Validate() const;
+
+  std::string ToString() const;
+};
+
+}  // namespace sherpa_onnx
+
+#endif  // SHERPA_ONNX_CSRC_OFFLINE_MEDASR_CTC_MODEL_CONFIG_H_