Populate ys_probs in the online NeMo transducer greedy search decoders

[josephomills]

Fixes #3181

What this PR does

The online NeMo transducer greedy search decoder never fills ys_probs, so streaming NeMo results always report "ys_probs": [] and applications cannot access per-token confidence. The offline path was fixed in #3105; this applies the same approach to the online decoders:

• after the argmax, apply LogSoftmax to the joiner output and push the selected token's log-probability alongside the token (only when a non-blank is emitted, so ys_probs.size() == tokens.size()).

LogSoftmax runs after the token has been selected, so decoding output is unchanged.

The same gap existed in the parakeet-unified online decoder; it gets the identical fix. The multilingual language-tag filter (FilterLanguageTags in online-recognizer-transducer-nemo-impl.h) already filters ys_probs index-aligned with tokens whenever the sizes match, so tag-filtered results stay aligned with no further changes.

Downstream context: FreeShow, an open-source church presentation app with an AI scripture feature under review (ChurchApps/FreeShow#3579), runs streaming Nemotron via sherpa-onnx-node and uses these per-token log-probs for confidence gating on live transcripts.

Changes

• sherpa-onnx/csrc/online-transducer-greedy-search-nemo-decoder.cc: LogSoftmax + r.ys_probs.push_back(log_prob) (mirrors the offline feat: Add ys_log_probs to NeMo transducer greedy search decoder #3105 fix; +1 include of math.h).
• sherpa-onnx/csrc/online-transducer-greedy-search-nemo-parakeet-unified-decoder.cc: identical fix.

Verification output

Streaming Nemotron 3.5 1120 ms int8 on macOS arm64, greedy_search, exercised through the node addon:

tokens: 48, ys_probs: 48        (lengths match; was ys_probs: 0 before)
all <= 0: true
min: -0.6439, max: -0.0001, mean: -0.1096
text unchanged vs pre-patch baseline: true
sample token/logprob pairs:
  " A"   -0.1219
  "ft"   -0.0010
  "er"   -0.1261

Notes

• Backward compatible: ys_probs already exists in the result struct and JSON — it just goes from always-empty to populated for these decoders.
• No temperature scaling, matching feat: Add ys_log_probs to NeMo transducer greedy search decoder #3105 (these decoders take no temperature_scale, unlike the standard online greedy decoder).
• ./scripts/check_style_cpplint.sh 1 passes. Built with the test-nodejs-addon-api.yaml cmake flags. The related CI workflows do not run on pull requests; verification above is local.

Summary by CodeRabbit

• Enhancements
  • Added log-probability scores for tokens emitted during online speech recognition.
  • Improved decoding output by associating each recognized token with its confidence-related score.

[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: 16c22147-aebd-4120-82b5-fdc4ad1c1faa

📥 Commits

Reviewing files that changed from the base of the PR and between d33f87c and 0927f7e.

📒 Files selected for processing (2)

• sherpa-onnx/csrc/online-transducer-greedy-search-nemo-decoder.cc
• sherpa-onnx/csrc/online-transducer-greedy-search-nemo-parakeet-unified-decoder.cc

Included review availability: Your plan provides up to 10 included reviews per hour; 7 remain after this review.

---

📝 Walkthrough

Walkthrough

Online NeMo greedy decoders now apply LogSoftmax to joiner logits and store the selected non-blank token's log probability in ys_probs.

Changes

NeMo decoder probability recording

Layer / File(s)	Summary
Record emitted-token log probabilities sherpa-onnx/csrc/online-transducer-greedy-search-nemo-decoder.cc, sherpa-onnx/csrc/online-transducer-greedy-search-nemo-parakeet-unified-decoder.cc	Both DecodeOne implementations apply LogSoftmax to joiner logits and append the selected token's log probability to ys_probs when a non-blank token is emitted.

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

Merge Risk: ⚪ Minimal · up to 0927f

This change populates per-token log-probabilities for two online decoders without changing decoded tokens; no actionable merge-blocking risk remains beyond normal checks and review.

Suggested reviewers: csukuangfj

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 2 functions across 2 files.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly and concisely describes the main change: populating ys_probs in the online NeMo transducer greedy search decoders.
Linked Issues check	✅ Passed	The changes satisfy issue #3181 by applying LogSoftmax to joiner logits and storing the selected token log-probability in ys_probs for the online NeMo decoder. The parakeet-unified decoder is also upd…
Out of Scope Changes check	✅ Passed	All changes are related to populating ys_probs for the specified online NeMo transducer greedy search decoders. No unrelated code changes are identified.

Full details: Linked Issues check

Explanation

The changes satisfy issue #3181 by applying LogSoftmax to joiner logits and storing the selected token log-probability in ys_probs for the online NeMo decoder. The parakeet-unified decoder is also updated as stated in the PR objectives.

✨ 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.}

Warning

⚠️ This pull request shows signs of AI-generated slop (redundant_comments). It has been flagged by CodeRabbit slop detection and should be reviewed carefully.

[csukuangfj]

Thank you for your contribution!