FireRedPunc: Bilingual Punctuation Predictor

• FireRedPunc is a bilingual punctuation restoration model that assigns punctuation tags to tokens using a BERT-style encoder.
• It operates as a token-level classifier within an integrated speech-processing pipeline, leveraging large, multi-domain training data.
• Benchmark results show an average F1 of 78.90, significantly surpassing FunASR-Punc in both Chinese and English evaluations.

FireRedPunc is the punctuation-prediction component of FireRedASR2S, introduced as part of an all-in-one speech-processing pipeline that also includes ASR, VAD, and LID. Its function is to convert an unpunctuated ASR transcript into a punctuated text sequence by assigning a punctuation label to each input token. In the reported system, FireRedPunc is designed for Chinese and English, uses a BERT-style encoder initialized from the LERT checkpoint, and is evaluated on internal multi-domain punctuation benchmarks where it attains an average F1 of 78.90, outperforming FunASR-Punc at 62.77 (Xu et al., 11 Mar 2026).

1. System role and task formulation

Within FireRedASR2S, FireRedPunc occupies the text-restoration stage: it receives an unpunctuated transcript and predicts which punctuation mark, if any, should follow each token. The report formulates this as token tagging over five labels: no-punctuation, comma, period, question mark, and exclamation mark. Concretely, each input token $t_i$ is assigned a label $$y_i \in \{\text{“〈no〉”}, \text{“，”}, \text{“。”}, \text{“？”}, \text{“！”}\}$$ indicating the punctuation mark to be inserted after that token (Xu et al., 11 Mar 2026).

This formulation makes FireRedPunc a sequence labeling model rather than a generative decoder. A plausible implication is that punctuation restoration is operationally decoupled from acoustic modeling in the reported implementation: the model operates over tokenized transcript text, not over speech frames. That distinction matters in industrial ASR pipelines because punctuation prediction can then be developed, trained, and evaluated as an independent text-side module.

2. Encoder architecture and classification head

FireRedPunc “adopts a BERT-style encoder” initialized from the LERT checkpoint of Cui et al. (2022). The report characterizes LERT as a standard Transformer encoder with 12 identical layers. Each layer consists of multi-head self-attention with $h = 12$ heads, model dimension $d_{\text{model}} = 768$, key/query/value dimension $d_k = d_v = 64$, and a feed-forward sublayer with inner size 3072. The model also uses absolute positional embeddings of size 768, together with Pre-LayerNorm and residual connections around both the self-attention and feed-forward sublayers (Xu et al., 11 Mar 2026).

The attention mechanism is restated in the report in standard Transformer form:

$$\mathrm{head}_i = \mathrm{Attention}(QW_i^Q,\;KW_i^K,\;VW_i^V), \quad \mathrm{MultiHead}(Q,K,V) = \mathrm{Concat}(\mathrm{head}_1,\dots,\mathrm{head}_{h})\,W^O,$$

with

$$\mathrm{Attention}(Q,K,V) = \mathrm{softmax}\Bigl(\tfrac{QK^\top}{\sqrt{d_k}}\Bigr)\,V.$$

On top of the encoder’s final hidden states $\{h_i\}$, FireRedPunc applies a token-wise linear projection and softmax to produce a distribution over the five punctuation classes for each token. The architecture is therefore conventional in its modeling interface: contextual encoding followed by per-token classification. What distinguishes the module in the report is not a novel classifier topology, but the combination of a pretrained LERT backbone, large-scale multi-domain training data, and bilingual punctuation restoration in Chinese and English.

3. Input representation and label alignment

During both training and inference, FireRedPunc uses the same tokenizer as the LERT encoder. The report specifies that this implies character-level segmentation for Chinese, where each Chinese character is one token, and WordPiece/BPE subword units for English, as in standard BERT (Xu et al., 11 Mar 2026).

The label space is aligned directly to this tokenization. Each token receives the label for the punctuation mark that should follow it in the reconstructed text. This is especially consequential in bilingual settings because the granularity of tokenization differs substantially across the two supported languages: Chinese is represented at the character level, whereas English is represented at the subword level. This suggests that the model’s notion of punctuation context is partly mediated by language-specific token boundaries inherited from LERT rather than by any punctuation-specific segmentation scheme.

The report does not describe any additional normalization such as lowercasing or truecasing. Instead, text is tokenized with the LERT vocabulary, and the token sequence is passed directly into the encoder.

4. Training data and optimization objective

FireRedPunc is trained on a large multi-domain text corpus comprising approximately 18.57 billion Chinese characters and 2.20 billion English words, drawn from news, conversational transcripts, subtitles, web text, and related sources, all annotated with “gold” punctuation. The scale and domain diversity are central characteristics of the reported training setup (Xu et al., 11 Mar 2026).

The training objective is standard token-level categorical cross-entropy over the five classes. If $\hat y_{i,c}$ is the predicted probability of class $c$ at token $$y_i \in \{\text{“〈no〉”}, \text{“，”}, \text{“。”}, \text{“？”}, \text{“！”}\}$$0 and $$y_i \in \{\text{“〈no〉”}, \text{“，”}, \text{“。”}, \text{“？”}, \text{“！”}\}$$1 is the one-hot gold label, the loss is

$$y_i \in \{\text{“〈no〉”}, \text{“，”}, \text{“。”}, \text{“？”}, \text{“！”}\}$$2

The report is explicit about several aspects of the setup and silent about others. It states that text is tokenized with the LERT vocabulary and that no additional normalization is reported. By contrast, it does not specify the optimizer, learning-rate schedule, batch size, or number of fine-tuning epochs. For researchers interested in reproducibility, this means that the reported training formulation is clear at the objective level but incomplete at the optimization-protocol level.

5. Evaluation protocol and reported benchmark results

FireRedPunc is evaluated on two internal multi-domain punctuation test sets: a Chinese set with 88,644 sentences and an English set with 28,641 sentences. The reported metrics are precision, recall, and F1, computed at each punctuation token. “Overall” F1 is defined as micro-averaged across all non-empty punctuation marks, excluding the rare exclamation mark (Xu et al., 11 Mar 2026).

The reported results are summarized below.

Test set	FireRedPunc (P / R / F1)	FunASR-Punc (P / R / F1)
Multi-domain Chinese	82.84 / 83.08 / 82.96	77.27 / 74.03 / 75.62
Multi-domain English	78.40 / 71.57 / 74.83	55.79 / 45.15 / 49.91
Average F1	78.90	62.77

These numbers indicate that the reported advantage over FunASR-Punc is present in both supported languages and is particularly large on the English benchmark. The Chinese result is more balanced in precision and recall, whereas the English result shows a larger recall deficit relative to precision. Because the paper does not provide a dedicated per-class breakdown or confusion matrix, the source of that asymmetry cannot be localized further from the published evidence alone.

6. Deployment characteristics, omissions, and future directions

The report does not provide explicit model-size, latency, or throughput numbers for FireRedPunc. Since it reuses a BERT-base-style encoder, one may infer a parameter count on the order of 100–120 million parameters. No streaming or low-latency variant is described (Xu et al., 11 Mar 2026).

This absence is operationally significant. FireRedASR2S as a whole includes modules with explicit deployment-oriented properties—for example, the VAD component is described as ultra-lightweight and supports streaming and non-streaming modes—whereas FireRedPunc is presented primarily through its architecture and benchmark accuracy. It should therefore not be assumed, on the basis of the report alone, that the punctuation module is optimized for online inference or incremental decoding.

The paper also does not include a dedicated error analysis for FireRedPunc. There is no confusion matrix, no per-class error breakdown, and no analysis of domain-specific failure modes. In the conclusion, the authors state that “future work will focus on further improving performance and expanding support for more languages.” The report further notes several possible enhancements, though not as detailed implementation proposals: incorporating prosodic or acoustic features alongside text to disambiguate punctuation in spoken inputs; leveraging relative-position embeddings or advanced attention mechanisms to capture long-range dependencies; and adapting larger pretrained LLMs or in-context learning (“prompting”) for more flexible punctuation insertions in novel domains. These are presented as possibilities rather than as completed extensions.

7. Position within punctuation restoration research

As reported, FireRedPunc is a bilingual punctuation restoration model that combines a standard Transformer-encoder formulation with large-scale multi-domain supervision and integration into an industrial speech pipeline. Its immediate research significance lies less in architectural novelty than in the way it instantiates punctuation prediction as a practical subsystem inside FireRedASR2S: token-level labeling, LERT-based contextual encoding, unified support for Chinese and English, and benchmarked comparison against FunASR-Punc (Xu et al., 11 Mar 2026).

The module also clarifies a recurrent point of confusion in speech-text post-processing research: punctuation restoration need not be solved inside the acoustic recognizer itself. FireRedPunc is reported as a distinct component that operates on the transcript after recognition. This suggests a modular design philosophy in which ASR, VAD, LID, and punctuation prediction can be improved independently while still contributing to an integrated production pipeline. At the same time, the reported future directions imply that purely text-based tagging may not exhaust the problem, especially where punctuation depends on prosody, long-range discourse structure, or transfer to previously unseen domains and languages.