paraformer_v2_community offline #2785

dengcunqin · 2026-01-17T10:14:29Z

This PR provides a code reproduction of Paraformer-v2: An Improved Non-Autoregressive Transformer for Noise-Robust Speech Recognition, as described in the paper. The implementation may not be fully consistent with the official code but aims to replicate the core concepts and methods presented in the paper.

gemini-code-assist · 2026-01-17T10:14:45Z

Summary of Changes

Hello @dengcunqin, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request introduces a complete offline implementation of the Paraformer-v2 model, a non-autoregressive transformer architecture tailored for noise-robust speech recognition. The core changes involve a specialized SANM-based decoder and a comprehensive model class that integrates CTC for alignment and prediction, enabling efficient and accurate speech-to-text conversion.

Highlights

New Decoder Architecture: Introduced DecoderLayerSANM and ParaformerSANMDecoder classes, specifically designed for the Paraformer-v2 model, extending BaseTransformerDecoder to handle self-attention and cross-attention mechanisms.
Paraformer-v2 Model Implementation: Added the main Paraformer class, which integrates the encoder, the new SANM decoder, and CTC components for end-to-end speech recognition, enabling a complete offline reproduction of the model.
CTC Alignment and Averaging Utilities: Implemented map_alignment_to_target_index, force_align, average_repeats_training, and average_repeats_inference methods to robustly handle CTC-based token alignment and aggregation for both training and inference processes.
Model Configuration File: Included a template.yaml file to define the model's configuration, specifying parameters for the encoder, decoder, frontend, specaug, and various training settings.

🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	`/gemini review`	Performs a code review for the current pull request in its current state.
Pull Request Summary	`/gemini summary`	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	`/gemini help`	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

gemini-code-assist

Code Review

This pull request introduces an offline implementation of Paraformer-v2. The code is generally well-structured, but I've identified a few issues that need attention. There's a critical bug in decoder.py that would lead to a runtime error, and a logic bug in handling decoder blocks which could affect model performance. Additionally, model.py has some inconsistencies in return types and unused variables. I've also pointed out a consistent typo (sanm_shfit instead of sanm_shift) in both the Python code and the YAML configuration, which should be corrected for better maintainability. My detailed comments and suggestions are provided below.

gemini-code-assist · 2026-01-17T10:16:31Z

funasr/models/paraformer_v2_community/decoder.py

+        memory_mask = myutils.sequence_mask(hlens, device=memory.device)[:, None, :]
+
+        tgt, tgt_mask, memory, memory_mask, _ = self.decoders[0](tgt, tgt_mask, memory, memory_mask)
+        attn_mat = self.model.decoders[1].get_attn_mat(tgt, tgt_mask, memory, memory_mask)


The code uses self.model.decoders, but self.model is not an attribute of the ParaformerSANMDecoder class. This will raise an AttributeError at runtime. It should probably be self.decoders.

Suggested change

attn_mat = self.model.decoders[1].get_attn_mat(tgt, tgt_mask, memory, memory_mask)

attn_mat = self.decoders[1].get_attn_mat(tgt, tgt_mask, memory, memory_mask)

gemini-code-assist · 2026-01-17T10:16:31Z

funasr/models/paraformer_v2_community/decoder.py

+                look_back=cache["decoder_chunk_look_back"],
+            )
+
+        if self.num_blocks - self.att_layer_num > 1:


The condition if self.num_blocks - self.att_layer_num > 1: seems incorrect. If self.num_blocks - self.att_layer_num is exactly 1, self.decoders2 would be initialized with one layer, but this block would be skipped. This is likely a bug. You should probably check if the value is greater than 0, or more robustly, check if self.decoders2 is not None.

Suggested change

if self.num_blocks - self.att_layer_num > 1:

if self.decoders2 is not None:

gemini-code-assist · 2026-01-17T10:16:31Z

funasr/models/paraformer_v2_community/decoder.py

+            )
+            new_cache.append(c_ret)
+
+        if self.num_blocks - self.att_layer_num > 1:


Similar to forward_chunk, the condition if self.num_blocks - self.att_layer_num > 1: seems incorrect and would cause the decoders2 block to be skipped if it contains exactly one layer. This should likely be if self.decoders2 is not None: to ensure the block is executed whenever it's initialized.

Suggested change

if self.num_blocks - self.att_layer_num > 1:

if self.decoders2 is not None:

gemini-code-assist · 2026-01-17T10:16:31Z

funasr/models/paraformer_v2_community/model.py

+            timestamps: List[Tuple[int, int]] -> [(start_frame, end_frame), ...]
+        """
+        if greedy_path.numel() == 0:
+            return torch.zeros((0, ctc_probs.size(1)), device=ctc_probs.device), []


This function has inconsistent return types. Here, it returns a tuple (torch.Tensor, []), but in other branches (lines 438, 440) it returns a single tensor. The caller at line 508 expects a single tensor. This will cause a TypeError if this branch is taken. The empty list seems to be related to timestamps which is mentioned in the docstring but is otherwise unused. You should probably remove the empty list from the return value to make it consistent and update the docstring to reflect the actual return value.

Suggested change

return torch.zeros((0, ctc_probs.size(1)), device=ctc_probs.device), []

return torch.zeros((0, ctc_probs.size(1)), device=ctc_probs.device)

gemini-code-assist · 2026-01-17T10:16:31Z

funasr/models/paraformer_v2_community/decoder.py

+        concat_after: bool = False,
+        att_layer_num: int = 6,
+        kernel_size: int = 21,
+        sanm_shfit: int = 0,


There's a typo in the parameter name sanm_shfit. It should be sanm_shift for better readability and consistency. This typo is also present in its usage within the __init__ method and in the template.yaml file.

Suggested change

sanm_shfit: int = 0,

sanm_shift: int = 0,

gemini-code-assist · 2026-01-17T10:16:31Z

funasr/models/paraformer_v2_community/decoder.py

+        if sanm_shfit is None:
+            sanm_shfit = (kernel_size - 1) // 2


There's a typo here, sanm_shfit should be sanm_shift.

Suggested change

if sanm_shfit is None:

sanm_shfit = (kernel_size - 1) // 2

if sanm_shift is None:

sanm_shift = (kernel_size - 1) // 2

gemini-code-assist · 2026-01-17T10:16:31Z

funasr/models/paraformer_v2_community/decoder.py

+            lambda lnum: DecoderLayerSANM(
+                attention_dim,
+                MultiHeadedAttentionSANMDecoder(
+                    attention_dim, self_attention_dropout_rate, kernel_size, sanm_shfit=sanm_shfit


There's a typo here, sanm_shfit should be sanm_shift.

Suggested change

attention_dim, self_attention_dropout_rate, kernel_size, sanm_shfit=sanm_shfit

attention_dim, self_attention_dropout_rate, kernel_size, sanm_shift=sanm_shift

gemini-code-assist · 2026-01-17T10:16:31Z

funasr/models/paraformer_v2_community/model.py

+            # Handling Noise/Silence (Empty Output)
+            if compressed_prob.size(0) == 0:
+                token_int = []
+                timestamp_list = []


The variable timestamp_list is initialized here but it is never used. This appears to be dead code and should be removed to improve clarity.

gemini-code-assist · 2026-01-17T10:16:31Z

funasr/models/paraformer_v2_community/template.yaml

+    pos_enc_class: SinusoidalPositionEncoder
+    normalize_before: true
+    kernel_size: 11
+    sanm_shfit: 0


There's a typo in the configuration key sanm_shfit. It should be sanm_shift to match the corrected parameter name in the model implementation.

sanm_shift: 0

gemini-code-assist · 2026-01-17T10:16:31Z

funasr/models/paraformer_v2_community/template.yaml

+    src_attention_dropout_rate: 0.1
+    att_layer_num: 16
+    kernel_size: 11
+    sanm_shfit: 0


There's a typo in the configuration key sanm_shfit. It should be sanm_shift.

sanm_shift: 0

gemini-code-assist bot reviewed Jan 17, 2026

View reviewed changes

Fix return statement for empty greedy path case

7f39f62

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

paraformer_v2_community offline #2785

paraformer_v2_community offline #2785

dengcunqin commented Jan 17, 2026

Uh oh!

gemini-code-assist bot commented Jan 17, 2026

Uh oh!

gemini-code-assist bot left a comment

Uh oh!

gemini-code-assist bot Jan 17, 2026

Uh oh!

gemini-code-assist bot Jan 17, 2026

Uh oh!

gemini-code-assist bot Jan 17, 2026

Uh oh!

gemini-code-assist bot Jan 17, 2026

Uh oh!

gemini-code-assist bot Jan 17, 2026

Uh oh!

gemini-code-assist bot Jan 17, 2026

Uh oh!

gemini-code-assist bot Jan 17, 2026

Uh oh!

gemini-code-assist bot Jan 17, 2026

Uh oh!

gemini-code-assist bot Jan 17, 2026

Uh oh!

gemini-code-assist bot Jan 17, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

	attn_mat = self.model.decoders[1].get_attn_mat(tgt, tgt_mask, memory, memory_mask)
	attn_mat = self.decoders[1].get_attn_mat(tgt, tgt_mask, memory, memory_mask)

	if self.num_blocks - self.att_layer_num > 1:
	if self.decoders2 is not None:

	return torch.zeros((0, ctc_probs.size(1)), device=ctc_probs.device), []
	return torch.zeros((0, ctc_probs.size(1)), device=ctc_probs.device)

	attention_dim, self_attention_dropout_rate, kernel_size, sanm_shfit=sanm_shfit
	attention_dim, self_attention_dropout_rate, kernel_size, sanm_shift=sanm_shift

paraformer_v2_community offline #2785

Are you sure you want to change the base?

paraformer_v2_community offline #2785

Conversation

dengcunqin commented Jan 17, 2026

Uh oh!

gemini-code-assist bot commented Jan 17, 2026

Summary of Changes

Highlights

Footnotes

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist bot Jan 17, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 17, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 17, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 17, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 17, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 17, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 17, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 17, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 17, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 17, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant