⚡ electrical engineering

Audio-Based Understanding of Audiobook Narration Appeal

This study presents the first systematic computational analysis demonstrating that vocal and acoustic features extracted from audiobook narration robustly predict listener appeal across different genres and titles, offering valuable data-driven insights for personalization and narrator casting.

Original authors: Shahar Elisha, Mariano Beguerisse-Díaz, Emmanouil Benetos

Published 2026-07-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Shahar Elisha, Mariano Beguerisse-Díaz, Emmanouil Benetos

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are walking into a massive library where every single book has been read aloud by a different person. Some readers sound like soothing grandmothers, others like energetic teenagers, and some sound like they are reading a grocery list. The paper you're asking about is essentially a scientific investigation into how the voice of the reader changes whether you actually enjoy the book or stop listening halfway through.

Here is the breakdown of their study using simple analogies:

1. The Big Question: Is the Reader as Important as the Story?

Think of an audiobook like a meal. The text is the recipe (the ingredients), and the narrator is the chef. You might have a great recipe (a famous book), but if the chef burns the food or serves it at the wrong temperature, you won't eat it.

The researchers wanted to know: Does the "chef's style" (the voice, speed, and tone) actually matter, even if the "recipe" (the story) is the same? They suspected that a great voice could make a boring story interesting, or a bad voice could ruin a masterpiece.

2. The Ingredients They Analyzed

To study this, they didn't just listen with their ears; they used computer "microscopes" to break down the audio into tiny, measurable pieces. They looked at:

The "Pace" (Tempo): How fast or slow the reader is talking.
The "Volume" (Energy): How loud or quiet the voice is.
The "Texture" (Spectral features): Is the voice breathy, raspy, or smooth?
The "Music" (Sound events): Are there sound effects or background music, or is it just a pure voice?

They gathered data from LibriVox, a giant public library of audiobooks read by volunteers. It's like a "community potluck" of stories rather than a professional restaurant, meaning the recording quality varies wildly—some sound like they were recorded in a soundproof studio, others like they were recorded in a noisy kitchen.

3. The Experiment: Measuring "Popularity"

Since they couldn't ask millions of people, "Did you like this?", they used a clever proxy: Page Views.

The Analogy: Imagine a book on a shelf. If people keep walking up to it, looking at it, and picking it up, it's popular. If it sits there gathering dust, it's not.
They counted how many times people viewed a specific audiobook recording. They adjusted this number based on how long the book had been on the shelf (so an old book doesn't get credit just for being there longer).

4. What They Found

Even with "noisy" data (volunteer recordings and simple view counts), they found some clear patterns:

The Voice Matters: The computer models proved that the sound of the voice alone can predict how popular a book is. It's not just about the story; the "chef's style" is a huge factor.
One Size Does Not Fit All: What works for a romance novel doesn't work for a history book.
- Analogy: A "breathy" voice might be like a soft, romantic candlelight dinner for a romance novel (people loved it). But that same breathy voice might be like serving a steak with a dessert spoon for a history book (people didn't like it).
- For history books, a voice with more "effort" or clarity (measured by something called the Hammarberg index) was preferred.
Speed is Good: Generally, readers who spoke with a bit more variation in their speed (not a robotic, flat pace) were more appealing. It's like a good storyteller who speeds up during an action scene and slows down for a sad moment.
The "Same Book" Test: They looked at books that had multiple recordings (e.g., Pride and Prejudice read by 10 different people). They found that the difference in popularity between these versions was almost as big as the difference between two completely different books. This proves that the narrator is just as important as the title.

5. The "Spotify" Check (The Secret Sauce)

The authors knew that "page views" on a public site isn't a perfect measure of enjoyment. So, they ran a smaller, secret test using data from Spotify (a music and audiobook streaming service).

Instead of just counting views, they looked at Return Rate: Did the user come back to listen to the book again within two weeks?
The Result: When they used this "return" metric, the connection between the voice and the listener's enjoyment became even stronger. It confirmed that the voice isn't just a first impression; it keeps people coming back.

6. The Bottom Line

The paper concludes that how a story is told is just as critical as the story itself.

If you are a listener, your favorite book might depend more on the narrator's voice than you think.
If you are a platform (like Spotify or Audible), you can't just recommend books based on the plot. You need to match the right voice to the right genre and the right listener.

What they did NOT claim:

They did not say they can now perfectly predict which specific book you will love.
They did not claim this works for professional narrators only (they used volunteers).
They did not say this applies to medical or clinical uses (like therapy).
They simply proved that voice characteristics are a measurable, powerful driver of whether people stick with an audiobook.

Technical Summary: Audio-Based Understanding of Audiobook Narration Appeal

Problem Statement
Narration is a central component of the audiobook listening experience, capable of elevating or undermining listener engagement. While platforms like Spotify and Audible face challenges in distinguishing between multiple recordings of the same title and understanding which narration styles appeal to specific genres or users, existing research lacks a systematic, large-scale computational analysis linking narration qualities to real-world consumption. Prior work has focused on small datasets, synthetic speech, or qualitative surveys, failing to address how acoustic features interact with genre and title across diverse recordings. Furthermore, user-item interaction data in audiobooks is sparse compared to music, as books are rarely replayed, necessitating sophisticated modeling of both content and narration style to improve recommendation and personalization systems.

Methodology
The authors conducted a computational study using a large, publicly available dataset from LibriVox, comprising 8,854 English audiobooks read by 1,206 narrators across 65 genres. To ensure reproducibility and transparency, the study utilized "view-rate" (page views normalized by days since publication) as a proxy for appeal, acknowledging its limitations regarding listening completion and duration bias.

The methodology involved three primary stages:

Feature Extraction: Audio recordings were segmented into 30-second intervals (up to 20 per book). The authors extracted 129 interpretable acoustic and vocal features using pre-trained models:
- eGeMAPSv02: 84 functional features derived from low-level descriptors (frequency, energy, spectral, tempo) via openSMILE.
- YAMNet: 34 features representing audio events (speech, music, sound effects, recording quality) aggregated from 521 classification scores.
- whisper-tiny: 11 features derived from transcripts, including word/syllable rates and duration.
Statistical Modeling:
- Global Modeling: A Generalized Linear Model (GLM) was fitted to assess the relationship between acoustic features and view-rate, controlling for multicollinearity via Variance Inflation Factor (VIF) pruning.
- Genre-Specific Modeling: Separate GLMs were trained per genre to identify how acoustic correlates of appeal vary across different categories.
- Intra-Title Modeling: A Linear Mixed-Effects (LME) model was employed with random intercepts for "book-groups" (different recordings of the same text) to disentangle title-driven variability from acoustic influences.
Predictive Modeling:
- Classification: View-rates were binned into quartiles to convert the task into classification. Four shallow classifiers (Logistic Regression, SVM, XGBoost, MLP) were trained to predict appeal quartiles using acoustic features, genres, or a combination.
- Ranking: To evaluate the ability to rank different narrations of the same title, boosted tree models (XGBRanker, LGBMRanker) were trained to predict within-title rankings based on acoustic features. Performance was measured using Kendall's rank correlation coefficient ( $\tau$ ).
Validation: To assess robustness, the authors repeated key analyses on a subset of 3,428 audiobooks hosted on Spotify, replacing the coarse view-rate metric with a granular "return-rate" (proportion of distinct users returning within 14 days).

Key Results

Statistical Association: Despite the coarse nature of the LibriVox data and variable recording quality, acoustic features alone explained a measurable portion of the variance in appeal (Global GLM pseudo- $R^2$ = 0.09). Thirty-one acoustic features showed statistically significant effects. Notably, variation in vocal shimmer (associated with breathiness) had a positive effect, while higher spectral flux (linked to perceived gender) had a negative effect.
Genre and Title Nuances: The influence of acoustic features varied significantly by genre. For instance, vocal shimmer was a strong positive predictor in Romance but insignificant in History, where the Hammarberg index (vocal effort/emotion) was the dominant factor.
Intra-Title Impact: The variation in appeal across different narrations of the same title was nearly as large as the variation across different titles (0.52 vs. 0.54), underscoring that narration style contributes to engagement at a scale comparable to content differences. The LME model significantly outperformed the global GLM (AIC difference of 210), confirming the necessity of modeling title biases.
Predictive Power: Classification models using acoustic features achieved accuracy up to 0.35 (compared to a random baseline of 0.25), outperforming genre-only models. Combining acoustic features with genres yielded the best performance. Ranking models demonstrated a statistically significant ability to order narrations of the same title better than random, with Kendall's $\tau$ reaching approximately 0.28 when using the Spotify return-rate metric.
Robustness: The findings were reinforced using the proprietary Spotify return-rate metric, which yielded higher model fit ( $R^2$ = 0.16) and stronger ranking correlations, suggesting that audio features are more strongly associated with engagement when controlling for exposure.

Significance and Claims
The authors claim this is the first systematic computational study linking narration qualities, genre, title, and large-scale real-world audiobook consumption. The study demonstrates that acoustic information alone has a robust association with audiobook appeal, even after accounting for title effects and using limited consumption data.

The paper posits that these findings highlight the potential of data-driven insights to improve audiobook personalization and narrator casting. By establishing that measurable vocal and acoustic properties contribute to listener engagement beyond content-level factors, the work provides a foundation for matching the right voice to the right book and the right book to the right listener. The authors remain modest regarding specific causal drivers due to feature correlations and dataset limitations, but they assert that the results offer converging evidence for the importance of paralinguistic features in the audiobook ecosystem.