Fix streaming paraformer feature frontend drift from its training config

[slipstr34m]

Problem

The offline paraformer configures the feature frontend the model was trained with (offline-recognizer-paraformer-impl.h:210-217):

void InitFeatConfig() {
  // Paraformer models assume input samples are in the range
  // [-32768, 32767], so we set normalize_samples to false
  config_.feat_config.normalize_samples = false;
  config_.feat_config.window_type = "hamming";
  config_.feat_config.high_freq = 0;
  config_.feat_config.snip_edges = true;
}

The online paraformer sets only normalize_samples, in both constructors, so streaming decoding runs on the fbank defaults (povey window, high_freq=-400, snip_edges=false) while the model's upstream config.yaml (modelscope speech_paraformer, online variant) specifies window: hamming. We noticed this while running the streaming and offline paraformer on the same audio: the streaming output was consistently worse, and the frontend difference turned out to be the reason. A user in #3511 found the same drift and worked around it with --high-freq=0 --snip-edges=true --window-type=hamming. Related: #3511.

Fix

Give the online impl the same private InitFeatConfig() and call it from both constructors. The gap dates to 6038e2a (#263), which copied the offline paraformer's then-current frontend handling; 25f0a10 (#1148) later upgraded only the offline impl.

Tested

New test_streaming_paraformer in test_online_recognizer.py decodes test_wavs/2.wav of sherpa-onnx-streaming-paraformer-bilingual-zh-en (greedy search, deterministic):

before: 这个是平繁的啊不认识接下来 frequently 苹繁的
after:  这个是频繁的啊不认识接下来 frequently 频繁的

The audio says 频繁 twice; the default frontend misrecognizes it both times. Individual tokens on other utterances can flip either way, but aggregate accuracy improves (about 10 errors fixed on a 4-minute zh recording, converging toward the offline paraformer's output on the same audio). The test fails on master without this change and passes with it; the full test_online_recognizer.py suite passes with no new skips. ./scripts/check_style_cpplint.sh passes.

Only touches online-recognizer-paraformer-impl.h and its Python test; the offline paraformer and all other online recognizers are untouched.

Summary by CodeRabbit

• Bug Fixes

  • Improved Paraformer streaming speech recognition feature configuration for more consistent audio processing.

• Tests

  • Added coverage for streaming Paraformer recognition with bilingual Chinese-English transcription.
  • Validates decoding results against an expected transcript.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 262ef5fd-860e-4745-994e-a7d8b32465ee

📥 Commits

Reviewing files that changed from the base of the PR and between 2744aa2 and 2e4c91c.

📒 Files selected for processing (2)

• sherpa-onnx/csrc/online-recognizer-paraformer-impl.h
• sherpa-onnx/python/tests/test_online_recognizer.py

---

📝 Walkthrough

Walkthrough

Paraformer feature extraction settings are centralized in InitFeatConfig(), and both constructors use it. A streaming Paraformer Python test now decodes a padded wave file and verifies the expected bilingual transcription.

Changes

Paraformer streaming recognition

Layer / File(s)	Summary
Centralized feature configuration sherpa-onnx/csrc/online-recognizer-paraformer-impl.h	Both constructors call InitFeatConfig(), which sets normalization, window type, high-frequency cutoff, and edge-snipping options.
Streaming recognition regression test sherpa-onnx/python/tests/test_online_recognizer.py	Adds a conditional streaming Paraformer test that feeds padded audio, decodes to completion, and asserts the bilingual transcript.

Estimated code review effort: 2 (Simple) | ~10 minutes

Suggested labels: size:M

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly matches the main change: aligning the streaming Paraformer frontend with its training configuration.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[csukuangfj]

Thank you for your contribution!