Honghua Chen | Chen Honghua

Speech and music exploit distinct intrinsic timescales of the sensorimotor system

Tue, 30 Jun 2026 00:00:00 +0000

Speech and music consistently differ in their acoustic rhythms, despite the cross-cultural diversity in their surface forms. Here, we investigate whether speech and music rhythms are rooted in distinct intrinsic timescales of the sensorimotor system, which are separately recruited to support individual communication and group synchronization, respectively. Corpus analysis revealed that the timescales dominating speech (4-8 Hz) and music rhythms (< 2 Hz) separately emerge in infant laughter, babbling and cries, and that both speech and song rhythms mature at about age three. The functional division between the two timescales is further probed through sensorimotor synchronization experiments, in which participants vocalize or tap to sound sequences presented at different rates. The rhythm produced by individuals is strongest between 4 and 8 Hz. In contrast, the produced rhythm is best synchronized among participants below 2 Hz. Collectively, these findings reveal two characteristic timescales in the human sensorimotor system, i.e., a faster (4-8 Hz) timescale reflecting resonance in individual production and a slower (<2 Hz) timescale that fosters interpersonal synchronization. The two timescales provide plausible biological basis for the rhythms of speech and music.

Efficient neural encoding of time intervals in speech and complex sound sequences

Mon, 01 Jun 2026 00:00:00 +0000

Encoding time intervals in complex sound faces dual challenges: it must be precise and cover a broad dynamic range. In speech, for example, a ten millisecond lengthening of a syllable can signal stress or phrasal boundaries, yet the syllable duration distribution is long tailed beyond 500 ms and has variable statistics across speakers. Here, we propose that the auditory cortex employs efficient coding to represent time intervals. When listeners heard syllable sequences drawn from different duration distributions, the magnetoencephalographic (MEG) response from the temporal cortex scaled with syllable duration, which is characterized using the interval response function. Crucially, this interval response function met predictions of efficient coding. Its intercept and slope adapted to the mean and variance of syllable duration, respectively, and it consistently exhibited a compressive nonlinearity that reduced response skewness, consistent with a maximum entropy code. A computational model that constantly updates the inference of duration distribution provided an algorithmic account of this efficient coding, and intracranial electroencephalogram (iEEG) data confirmed the same principles during natural speech comprehension. Together, our findings reveal an efficient neural mechanism that supports precise encoding of highly variable time intervals in complex sound sequences.

Low-frequency cortical activity reflects context-dependent parsing of word sequences

Thu, 08 May 2025 00:00:00 +0000

During speech listening, it has been hypothesized that the brain builds representations of linguistic structures like sentences, which are tracked by neural activity entrained to the rhythm of these structures. Alternatively, others proposed that these sentence-tracking neural activities may reflect the predictability or syntactic properties of individual words. Here, to disentangle the neural responses to sentences and words, we design word sequences that are parsed into different sentences in different contexts. By analyzing neural activity recorded by magnetoencephalography, we find that low-frequency neural activity strongly depends on context—the difference between MEG responses to the same word sequence in two contexts yields a low-frequency signal, which precisely tracks sentences. The predictability and syntactic properties of words can partly explain the neural response in each context but not the difference between contexts. In summary, low-frequency neural activity encodes sentences and can reliably reflect how same-word sequences are parsed in different contexts.

Normalization and Adaptation in Neuroscience

Fri, 04 Apr 2025 00:00:00 +0000

Normalization as a canonical computation

Projects

Sun, 19 May 2024 00:00:00 +0000

Original speech and its echo are segregated and separately processed in the human brain

Thu, 15 Feb 2024 00:00:00 +0000

Speech recognition crucially relies on slow temporal modulations (<16 Hz) in speech. Recent studies, however, have demonstrated that the long-delay echoes, which are common during online conferencing, can eliminate crucial temporal modulations in speech but do not affect speech intelligibility. Here, we investigated the underlying neural mechanisms. MEG experiments demonstrated that cortical activity can effectively track the temporal modulations eliminated by an echo, which cannot be fully explained by basic neural adaptation mechanisms. Furthermore, cortical responses to echoic speech can be better explained by a model that segregates speech from its echo than by a model that encodes echoic speech as a whole. The speech segregation effect was observed even when attention was diverted but would disappear when segregation cues, i.e., speech fine structure, were removed. These results strongly suggested that, through mechanisms such as stream segregation, the auditory system can build an echo-insensitive representation of speech envelope, which can support reliable speech recognition.

Probing the “Creativity” of Large Language Models: Can models produce divergent semantic association?

Fri, 08 Dec 2023 00:00:00 +0000

Large language models possess remarkable capacity for processing language, but it remains unclear whether these models can further generate creative content. The present study aims to investigate the creative thinking of large language models through a cognitive perspective. We utilize the divergent association task (DAT), an objective measurement of creativity that asks models to generate unrelated words and calculates the semantic distance between them. We compare the results across different models and decoding strategies. Our findings indicate that (1) When using the greedy search strategy, GPT-4 outperforms 96% of humans, while GPT-3.5-turbo exceeds the average human level. (2) Stochastic sampling and temperature scaling are effective to obtain higher DAT scores for models except GPT-4, but face a trade-off between creativity and stability. These results imply that advanced large language models have divergent semantic associations, which is a fundamental process underlying creativity.

Experience

Tue, 24 Oct 2023 00:00:00 +0000

Learn JavaScript

Tue, 24 Oct 2023 00:00:00 +0000

Hugo Blox Builder is designed to give technical content creators a seamless experience. You can focus on the content and the Hugo Blox Builder which this template is built upon handles the rest.

Embed videos, podcasts, code, LaTeX math, and even test students!

On this page, you’ll find some examples of the types of technical content that can be rendered with Hugo Blox.

Video

Teach your course by sharing videos with your students. Choose from one of the following approaches:

Youtube:

{{< youtube w7Ft2ymGmfc >}}

Bilibili:

{{< bilibili id="BV1WV4y1r7DF" >}}

Video file

Videos may be added to a page by either placing them in your assets/media/ media library or in your page’s folder, and then embedding them with the video shortcode:

{{< video src="my_video.mp4" controls="yes" >}}

Podcast

You can add a podcast or music to a page by placing the MP3 file in the page’s folder or the media library folder and then embedding the audio on your page with the audio shortcode:

{{< audio src="ambient-piano.mp3" >}}

Try it out:

Test students

Provide a simple yet fun self-assessment by revealing the solutions to challenges with the spoiler shortcode:

{{< spoiler text="👉 Click to view the solution" >}}
You found me!
{{< /spoiler >}}

renders as

👉 Click to view the solution

You found me 🎉

Math

Hugo Blox Builder supports a Markdown extension for $\LaTeX$ math. You can enable this feature by toggling the math option in your config/_default/params.yaml file.

To render inline or block math, wrap your LaTeX math with {{< math >}}$...${{< /math >}} or {{< math >}}$$...$${{< /math >}}, respectively.

We wrap the LaTeX math in the Hugo Blox math shortcode to prevent Hugo rendering our math as Markdown.

Example math block:

{{< math >}}
$$
\gamma_{n} = \frac{ \left | \left (\mathbf x_{n} - \mathbf x_{n-1} \right )^T \left [\nabla F (\mathbf x_{n}) - \nabla F (\mathbf x_{n-1}) \right ] \right |}{\left \|\nabla F(\mathbf{x}_{n}) - \nabla F(\mathbf{x}_{n-1}) \right \|^2}
$$
{{< /math >}}

renders as

$$\gamma_{n} = \frac{ \left | \left (\mathbf x_{n} - \mathbf x_{n-1} \right )^T \left [\nabla F (\mathbf x_{n}) - \nabla F (\mathbf x_{n-1}) \right ] \right |}{\left \|\nabla F(\mathbf{x}_{n}) - \nabla F(\mathbf{x}_{n-1}) \right \|^2}$$

Example inline math {{< math >}}$\nabla F(\mathbf{x}_{n})${{< /math >}} renders as $\nabla F(\mathbf{x}_{n})$ .

Example multi-line math using the math linebreak (\\):

{{< math >}}
$$f(k;p_{0}^{*}) = \begin{cases}p_{0}^{*} & \text{if }k=1, \\
1-p_{0}^{*} & \text{if }k=0.\end{cases}$$
{{< /math >}}

renders as

$$ f(k;p_{0}^{*}) = \begin{cases}p_{0}^{*} & \text{if }k=1, \\ 1-p_{0}^{*} & \text{if }k=0.\end{cases} $$

Code

Hugo Blox Builder utilises Hugo’s Markdown extension for highlighting code syntax. The code theme can be selected in the config/_default/params.yaml file.

```python
import pandas as pd
data = pd.read_csv("data.csv")
data.head()
```

renders as

import pandas as pd
data = pd.read_csv("data.csv")
data.head()

Inline Images

{{< icon name="python" >}} Python

renders as

Python

Learn Python

Tue, 24 Oct 2023 00:00:00 +0000

Hugo Blox Builder is designed to give technical content creators a seamless experience. You can focus on the content and the Hugo Blox Builder which this template is built upon handles the rest.

Embed videos, podcasts, code, LaTeX math, and even test students!

On this page, you’ll find some examples of the types of technical content that can be rendered with Hugo Blox.

Video

Teach your course by sharing videos with your students. Choose from one of the following approaches:

Youtube:

{{< youtube w7Ft2ymGmfc >}}

Bilibili:

{{< bilibili id="BV1WV4y1r7DF" >}}

Video file

Videos may be added to a page by either placing them in your assets/media/ media library or in your page’s folder, and then embedding them with the video shortcode:

{{< video src="my_video.mp4" controls="yes" >}}

Podcast

You can add a podcast or music to a page by placing the MP3 file in the page’s folder or the media library folder and then embedding the audio on your page with the audio shortcode:

{{< audio src="ambient-piano.mp3" >}}

Try it out:

Test students

Provide a simple yet fun self-assessment by revealing the solutions to challenges with the spoiler shortcode:

{{< spoiler text="👉 Click to view the solution" >}}
You found me!
{{< /spoiler >}}

renders as

👉 Click to view the solution

You found me 🎉

Math

Hugo Blox Builder supports a Markdown extension for $\LaTeX$ math. You can enable this feature by toggling the math option in your config/_default/params.yaml file.

To render inline or block math, wrap your LaTeX math with {{< math >}}$...${{< /math >}} or {{< math >}}$$...$${{< /math >}}, respectively.

We wrap the LaTeX math in the Hugo Blox math shortcode to prevent Hugo rendering our math as Markdown.

Example math block:

{{< math >}}
$$
\gamma_{n} = \frac{ \left | \left (\mathbf x_{n} - \mathbf x_{n-1} \right )^T \left [\nabla F (\mathbf x_{n}) - \nabla F (\mathbf x_{n-1}) \right ] \right |}{\left \|\nabla F(\mathbf{x}_{n}) - \nabla F(\mathbf{x}_{n-1}) \right \|^2}
$$
{{< /math >}}

renders as

Example inline math {{< math >}}$\nabla F(\mathbf{x}_{n})${{< /math >}} renders as $\nabla F(\mathbf{x}_{n})$ .

Example multi-line math using the math linebreak (\\):

{{< math >}}
$$f(k;p_{0}^{*}) = \begin{cases}p_{0}^{*} & \text{if }k=1, \\
1-p_{0}^{*} & \text{if }k=0.\end{cases}$$
{{< /math >}}

renders as

$$ f(k;p_{0}^{*}) = \begin{cases}p_{0}^{*} & \text{if }k=1, \\ 1-p_{0}^{*} & \text{if }k=0.\end{cases} $$

Code

Hugo Blox Builder utilises Hugo’s Markdown extension for highlighting code syntax. The code theme can be selected in the config/_default/params.yaml file.

```python
import pandas as pd
data = pd.read_csv("data.csv")
data.head()
```

renders as

import pandas as pd
data = pd.read_csv("data.csv")
data.head()

Inline Images

{{< icon name="python" >}} Python

renders as

Python

AI for Psycholinguistics

Sat, 10 Jun 2023 00:00:00 +0000

Recently, I participated in the organization of a Chinese workshop about NLP at Chinese Psycholinguistic Society in Guangzhou in 2023 (with Prof. Nai Ding, Dr. Jiajie Zou, and Wei Liu). It provides a psycholinguistic research tool based on artificial intelligence language models in natural language processing, which is convenient to extract language features and carry out language-related tasks.

Creativity of Large Language Models

Mon, 03 Apr 2023 00:00:00 +0000

“Creativity is the defeat of habit by originality.” –Arthur Koestler

Human progress stems from continuous innovation in technology, culture, and the arts. This creativity is one of the defining traits of what we consider intelligence. In language, creative expressions—such as metaphors, humor, and more—play a crucial role in communication and linguistic evolution. If humans can possess creativity, can language models like GPT-4 also exhibit it?

At first glance, this seems unlikely: the essence of creativity lies in the ability to produce novel and applicable ideas, yet language models are trained on existing corpora—essentially fitting the distribution of human language that already exists. How, then, could such models generate anything truly original?

However, current usage suggests that many people are indeed employing GPT-4 for creative tasks. For instance, mathematician Terence Tao has noted that GPT-4 assists him in solving problems and proving mathematical theorems. Some researchers have found that GPT-4 can produce literary works at a human level and even propose scientific hypotheses. If models can indeed perform creative work, how should we measure their creativity concretely?

Recently, we has made some attempts to address this question 1. This study developed a creativity evaluation for language models based on the Divergent Association Test (DAT). In psychology, researchers have devised numerous creativity assessments, but many of these tests may not be reliable when applied to models, as the test items might have leaked into the training data. If creativity is assessed purely based on output, models could also generate content by retrieving from their training corpora. These confounding factors make it difficult to discern a model’s true creative capacity. However, some evaluations approach creativity from a more cognitive perspective, measuring it through semantic retrieval flexibility—DAT being one such method [2].

The DAT paradigm

The DAT test requires generating 10 nouns that are as semantically unrelated as possible:

“Generate 10 nouns that are as unrelated to each other as possible.”

After obtaining the 10 words, the semantic distance between each pair is calculated and averaged to produce the final DAT score, which ranges from 0 to 200. A higher DAT score indicates stronger divergent associative ability. This test demands the ability to break away from the everyday distribution of language and produce distant semantic associations—a significant challenge for language models. Surprisingly, we found that while smaller models indeed fall short of human performance, the most advanced models today have reached or even surpassed the average human level. For example, GPT-4 outperformed 96.1% of humans.

We also experimented with different generation methods: greedy search (selecting the highest-probability word) and top-p sampling (drawing from the top-p probable words). Sampling improved DAT scores for weaker models, but the gains were limited.

The DAT score of LLMs

But are the models truly that impressive? We identified an issue with the DAT test: the measurement of semantic distance relies on word embeddings, which are influenced by word frequency. Plotting word frequency against DAT scores revealed that lower-frequency words correlate with higher DAT scores. GPT-4 and GPT-3.5’s difference mainly lies in GPT-4’s tendency to generate lower-frequency words. After controlling for word frequency, both models still scored above average but only surpassed 61% (GPT-4) and 75% (GPT-3.5) of humans. Notably, previous DAT studies on humans did not regress out word frequency, yet the results still showed high validity and correlation with other creativity tests. This suggests that generating low-frequency words may itself be linked to creativity.

The DAT score regressed by word frequency

The study further explored the models’ generation strategies and the validity of the DAT paradigm (details omitted here). In summary, we proposed a semantic network-based evaluation for large language models’ creativity, finding that state-of-the-art models perform above the human average on DAT tasks. Given the long-standing debates around defining and measuring creativity, our study only examines it from the perspective of semantic network flexibility. It does not conclusively prove that models possess creativity—especially specialized (Pro-C) or eminent (Big-C) creativity seen in only a few individuals. Nevertheless, such flexibility, both behaviorally and neurologically, is closely tied to creativity [3]. Moreover, the fact that a model trained to fit language distributions can step beyond them—following DAT’s instructions to generate unrelated words—is a delightful discovery.

References

Chen, H., & Ding, N. (2023). Probing the “Creativity” of Large Language Models: Can models produce divergent semantic association? Findings of EMNLP. https://aclanthology.org/2023.findings-emnlp.858/
Olson, J. A., Nahas, J., Chmoulevitch, D., Cropper, S. J., & Webb, M. E. (2021). Naming unrelated words predicts creativity. PNAS, 118(25), e2022340118. https://doi.org/10.1073/pnas.2022340118
Beaty, R. E., Benedek, M., Silvia, P. J., & Schacter, D. L. (2016). Creative Cognition and Brain Network Dynamics. Trends in Cognitive Sciences, 20(2), 87–95. https://doi.org/10.1016/j.tics.2015.10.004

Honghua Chen | Chen Honghua

Speech and music exploit distinct intrinsic timescales of the sensorimotor system

Efficient neural encoding of time intervals in speech and complex sound sequences

Low-frequency cortical activity reflects context-dependent parsing of word sequences

Normalization and Adaptation in Neuroscience

Projects

Original speech and its echo are segregated and separately processed in the human brain

Probing the “Creativity” of Large Language Models: Can models produce divergent semantic association?

Experience

Learn JavaScript

Video

Podcast

Test students

Math

Code

Inline Images

Did you find this page helpful? Consider sharing it 🙌

Learn Python

Video

Podcast

Test students

Math

Code

Inline Images

Did you find this page helpful? Consider sharing it 🙌

AI for Psycholinguistics

Creativity of Large Language Models

References