From close listening to distant listening: building tools for speech-music discrimination in Danish P3 radio

From close listening to distant listening: developing tools for speech-music discrimination in Danish music radio

Digitization has transformed the landscape of music radio. Music streaming services such as Spotify and iTunes have largely outperformed traditional playlist-based radio, and the global rollout of software-generated playlists in public service stations during the 1990s has displaced the passionate radio host. Yet digitization has also opened new avenues for radio research. In Denmark, nearly all radio programming dating back to 1989 has been digitized, making it possible to systematically investigate historical shifts in radio content. This study explores the question: how has the distribution of music and talk on the Danish Broadcasting Corporation’s radio channel P3 evolved from 1989 to 2019?

We compare a qualitative case study with a large-scale analysis that covers more than 65,000 hours of radio. Methodologically, moving from close listening to a few programs to distant listening across thousands of hours enables us to assess the methods, results, strengths, and limitations of both approaches. Earlier research has shown that Convolutional Neural Networks (CNNs) trained to recognize images of audio spectrograms outperform other methods, like Support Vector Machines (SVMs). The large-scale study presented here demonstrates that this CNN approach generalizes well — even without fine-tuning — to classifying speech and music in Danish radio, achieving an overall accuracy of 98%.

Introduction

In the collaborative research project “A Century of Radio and Music in Denmark,” funded by the Independent Research Fund Denmark (2013–2018), eleven scholars investigated the current state and historical development of Danish public music radio. The project revealed how digitization has reshaped music radio in several ways. Competition from streaming platforms such as Spotify and iTunes challenges traditional playlist radio, while the adoption of software-generated playlists in public service stations during the 1990s pushed the passionate radio host aside. Meanwhile, talk radio has gained popularity alongside on-demand podcasts. As music radio channels lose listeners, governments and public discussions worldwide debate whether music radio still serves a public-service role worthy of public funding.

As part of this project, Have tested the hypothesis that Danish public service music radio shifted from prioritizing recorded music toward emphasizing spoken content — chattering hosts, news, and so on. This hypothesis was examined qualitatively through five case studies of the most popular morning music show on P3 between 1989 and 2016. P3 is a nationwide public service music channel in Denmark, comparable to Sweden’s P3, Norway’s P3, or the BBC’s Radio 1. In the present article, we investigate the same hypothesis quantitatively by expanding the dataset from five programs to 65,000 hours of P3 broadcasts.

Digitization is not only changing radio itself but also creating new opportunities for radio research. Archives worldwide are being digitized as part of cultural heritage preservation, and some now permit researcher access. Even as archive policies become more accommodating, tools are still needed to navigate the immense volume of audio material. Access to digital archives has transformed media studies, especially by enabling scholars to listen deeply to historical recordings and investigate how radio content and expression have changed. Humanities scholars possess strong skills in deep hermeneutic and aesthetic content analysis, but digital archives and improved metadata also open the door to large-scale analysis, allowing us to listen both closely and at a distance.

The terms “close listening” and “distant listening” draw inspiration from Franco Moretti’s concept of distant reading. Here, close listening refers to human sensorial listening, manual annotation, and hermeneutic interpretation of audio — a method that allows deep, detailed engagement but works best with small amounts of material. Distant listening, by contrast, uses computational models to do the “listening” when data is too vast for human processing. The strength of distant listening is its breadth, enabling analysis of huge archives using simple categories like music and speech.

In sound studies, “deep listening” has been used to describe concentrated contemplative listening, while Tanya Clement has introduced “distant listening” as a methodological term in digital humanities for large-scale machine processing of audio data. Large-scale audio analysis remains rare in digital humanities compared to text and image analysis. Most existing audio work deals with recordings of text in literary studies. In musicology, Music Information Retrieval (MIR) — part of the broader field of Audio Content Analysis — has grown in recent years, but large-scale analysis in media studies is still uncommon.

This study is based on the Danish digital radio archive and infrastructure LARM.fm, which provides access to nearly two million digitized Danish radio programs. No tool for large-scale analysis of this archive had been developed previously, so one ambition of this project is to demonstrate a viable approach. The article therefore has two aims: (1) to describe the methodological process and challenges of building a model for large-scale speech-music discrimination that answers the research question — how did the distribution of music and talk on P3 develop between 1989 and 2019? — and (2) to critically compare the methods, results, strengths, and shortcomings of the qualitative case study and the large-scale analysis.

We begin by presenting the LARM.fm archive, which serves as the basis for both studies. Next, we briefly summarize the analog, hand-annotated case study first presented in “A Lost Link between Music and Hosts: The Development of a Morning Music-Radio Show.” We will refer to this as the “case study.” Then we introduce the large-scale study in terms of its methodology, challenges, and results. In the following section, we compare the two studies, weighing the trade-offs between close listening and distant listening, before concluding.

The LARM.fm archive: a digital resource and infrastructure

LARM.fm was originally developed by the LARM Audio Research Archive project (2010–2014), a collaboration among ten research and cultural institutions funded by the Danish Ministry of Higher Education and Science. Since 2015 it has been part of the national DIGHUMLAB project. The platform was later rebuilt on HTML5 and expanded to include television content.

Denmark has had legal deposit of all broadcast material to the Royal Danish Library since 1987, and this material is included in the Danish Media Collection. From 2005 onward, the material has been born-digital, while nearly all analogue programs from 1989 to 2005 have been digitized. Legal restrictions mean the Media Collection is available to the public only via on-site computers at libraries. However, an agreement between the Danish Agency for Culture and copyright holders allows university faculty and students to stream — though not download — material through the library’s Mediestream platform or LARM.fm.

Unlike Mediestream, LARM.fm also includes older material from the archives of the Danish Broadcasting Corporation (DR), some of which dates back to 1925, the year national radio launched in Denmark. This older material is incomplete and consists mostly of text documents, as many programs were not saved or have not yet been digitized. Beyond access, LARM.fm offers search tools, options for organizing material, and annotation and sharing features.

The collection grows steadily as new Danish radio and television programs are broadcast. In November 2019, it contained over three million radio and TV programs and more than 200,000 OCR-scanned PDF files.

The interface of LARM.fm organizes material by type: TV programs, Radio programs, Radio news reports, Radio news manuscripts, Program guides, and Radio news guides. Radio and TV programs are available for streaming; the other types consist of OCR-scanned PDF documents — some handwritten, others typed or printed.

The oldest radio program in the archive dates from May 22, 1931, lasting just over five minutes. There are ten or fewer programs from the first five years of that decade and between 35 and 61 from the remaining years of the 1930s.

LARM.fm has expanded Danish radio research toward deep, detailed content analysis based on primary sources — a kind of analysis that was previously lacking not only in Danish research but internationally within media studies, where radio studies often focused on institutional and media-systemic questions. Access to archives and digital tools like LARM.fm enables deeply contextualized synchronic analysis and, above all, supports historical diachronic study of longer periods, offering unique ways to combine qualitative and quantitative methods. The work presented here is one such example, coupling an existing qualitative study with a large-scale analysis, while also fostering discussion of methodological strengths and challenges in digital humanities.

The case study

The close-listening case study examined the development of music-to-speech balance in the most popular P3 music program, Go' Morgen P3. The program launched in November 1989, airing music from 6 to 9 a.m. on weekdays. Apart from changing hosts, the show remained relatively stable in format over the decades.

The case study aimed to test the hypothesis that between 1989 and 2016, Go' Morgen P3 transitioned from music as its primary content — both qualitatively and quantitatively — toward an increasing focus on speech and entertaining hosts. Have discussed this hypothesis in light of the introduction of computerized playlists and rotation systems in Danish public service radio in the 1990s, as well as growing competition from commercial and digital radio stations and, particularly, music streaming services. Here, we present only the methods and results relevant to the large-scale study.

Have listened closely to five Go' Morgen P3 episodes aired between 6 and 9 a.m. on Wednesdays near November 1, spanning 25 years. The programs were first roughly annotated using three categories: music, speech, and other. “Speech” included only host and interviewee speech; “other” covered news, sports, weather forecasts, jingles, and channel advertisements. The annotation was validated and fine-tuned by student assistants. Choosing the same weekday and approximate date across years helped make yearly comparisons more valid, while Wednesdays around November 1 avoid bank or festive holidays. The goal was an even distribution over time, but limited availability of older programs in LARM.fm resulted in episodes from 1992, 1998, 2006, 2011, and 2016.

To check that the 1992 episode represented the early 1990s and was not affected by DR’s institutional restructuring or channel reformatting, two additional episodes from 1990 and 1991 were added for comparison. The distribution among categories was consistent across these three years, confirming that 1992 was typical. The 2006 episode, unfortunately, suffered technical problems that caused extra “rescue music” in the first 15 minutes, inflating music time. This issue was discovered too late, so the episode remained in the sample.

The annotations revealed both stable and shifting elements. While the basic format stayed fairly constant, there were significant changes in content and in the relationship between music and hosts. Over 25 years, the program went from one host speaking between tracks to music filling the gaps between a group of three or four hosts. Table 1 shows a slight decrease in recorded music and, overall, more host speech across the years. The increase in speech did not come at the expense of music as originally expected, but rather at the expense of the “other” category.

Date/Category	4 Nov. 1992	28 Oct. 1998	25 Oct. 2006	2 Nov. 2011	2 Nov. 2016
Speech	10.28 min	26.48	46.3	55.52	51.5
Music	116.52	92.5	100.32	91.45	101.95
Other	42.22	51.07	28.63	26.15	22.5
Table 1. Go' Morgen P3 program content in minutes by category.

Nevertheless, close listening showed that the emphasis on music within the talk gradually declined. The time spent introducing music shrank significantly, and the connection between music and host weakened, disappearing entirely by 2016. In the early 1990s, the host chose the music and scripted much of their speech around it — practically everything said between tracks concerned the music. After playlists were introduced, this link ended. In 2011, the host still said “Chris Brown, With You,” but by 2016, such introductions vanished.

The same content elements reappear in all five episodes: recorded music, time announcements, and channel or program adverts. However, new entertaining elements unrelated to music — such as satire, quizzes, daily news, and current events — gradually entered the program. Combined with the hosts’ indifference to the music, this development pushed music further into the background.

Based on the case study, Go' Morgen P3 was concluded to have evolved from a program strongly centered on music into a lighter, entertainment-oriented format with chatter and soft news. This shifted the program from a clear music-radio format — where recorded music was central and essential — to a more ambiguous soft-news format, where music plays various roles, including underscore music and jingles crucial to the flow. These observations, made possible through close listening combined with knowledge of Danish media politics and institutional changes at DR, inform the large-scale analyses presented next. Instead of relying on a small, hand-annotated sample, this new study uses thousands of radio programs to test the qualitative findings on data from LARM.fm. This study marks the first large-scale audio analysis conducted on the Danish Media Collection.

The large-scale study

In this study, we extend the earlier study by zooming out: a computer does the “listening” and “counting” — not on five or seven episodes but on a sample of all P3 morning programming from 1989 to 2019, including the Go' Morgen P3 slot. These research questions guided the work:

• How did the distribution of music and talk develop between 1989 and 2019? (classification)
• How long are sequences of speech and music?

The corpus includes all radio programs broadcast on P3 between 6 a.m. and 12 p.m. from 1989 to 2019 available in the archive.

Analysts then applied the model developed by Papakostas and Giannakopoulos (2018) for Speech-Music discrimination. Their approach uses a convolutional neural network (CNN) for image-based classification on spectrograms, demonstrating state-of-the-art results. The model was built using transfer-learning from a CNN previously trained on a subset of ImageNet [Deng et al. 2009]. In this process, each audio file is divided into overlapping mid-term segments of 2.4 seconds with a 1-second step. A spectrogram is generated from each segment, and the CNN determines whether it contains speech or music. A median filter subsequently smooths the signal, removing, for example, unlikely 1-second speech segments within music.

During testing, this image classification method achieved accuracy of about 93-98% on 11 hours of radio stream [Papakostas and Giannakopoulos 2018], outperforming state-of-the-art audio-based classifiers, which typically reach 85-94%. For details on training, performance, and audio-to-spectrogram preprocessing, readers should consult the original study.

As shown in Table 2 and Figure 5, these performance levels generalized effectively to Danish when compared with our coded dataset. For an audio-based comparison, we employed the support vector machine (SVM) by Giannakopoulos (2015).

These performance metrics only include pure speech and music categories, which underrepresents the mixed content in music radio programs, especially after 2006 (see Figure 4). At that point, the proportion of the mixed "speech and music" category increased noticeably. By placing "Speech and music" into the "speech" category and "jingles" into "music", the CNN achieved an accuracy of 0.96 (95% CI: 0.954, 957). This calculation includes categories the CNN cannot predict, like noise and silence. Therefore, this figure likely represents the expected performance across all categories, assuming jingles count as music and "speech and music" as speech.

Such performance enables large-scale analysis of speech-music relationships. While our work focuses on radio classification, the approach applies broadly to other audio media and recordings such as podcasts, audiobooks, and music performances. Predictors can also be altered to include, for instance, gender or mood, and outcomes could be expanded to multiple categories like speech, music, and jingles.

The analysis results appear in the following four figures, discussed further in the subsequent section.

Figure 6 shows the model's predicted proportion of speech over time, with shaded areas indicating 95% confidence intervals. Data has been smoothed with a Kolmogorov-Zurbenko filter using a window size of 4 for clarity. Significant changes in speech proportion appear around 1992, 2001, and again in 2005, plus an upward trend from 2016 to 2019.

In addition to speech-music discrimination, we examined how speech and music segment lengths changed over the years. This interest stemmed from case study findings that musical track lengths have become standardized, particularly after the introduction of playlists, rotation systems, and musical clock schemes [Russo 2013].

Figure 7 tracks average segment length in seconds over time, using a Kolmogorov-Zurbenko filter with window size 6. Confidence intervals are shown as shading. Note that any interruption of a segment—say, a radio host cutting into music—halves that segment, so results should be interpreted accordingly.

Because Figure 7 displays mean length in seconds, the distribution is Poisson-like for speech and resembles a bimodal Poisson for music (see Figures 8 and 9). The average music length increased in 1992, while both music and speech lengths declined from 2011 onward, suggesting shorter segments and more frequent switching. Speech length rose on average from 1992 until 1999.

To clarify the trends in Figure 7, we grouped data into seven-year periods for separate comparisons of speech and music segments.

Figures 8 and 9 highlight a first peak of very short segments (10-20 seconds) and a second peak around 200 seconds, especially pronounced for music. These patterns reflect overlapping host talk in early years and, for the later peak, a standardization of musical track lengths following the 1992 restructuring. This illustrates what Have (2018) describes as the radiotization of music—radio's logic shapes the creative output of musicians. Compose tracks close to 3 minutes 20 seconds to fit the musical clock structure, and you get played. The clock format, as Hendy (2000, p.94ff) explains—also termed the format clock by Michelsen et al. (2018a, p.192f) and Russo (2013)—divides each hour into alternating speech and music segments. A duration just over 3 minutes per segment proved optimal for retaining listeners, balancing familiarity with novelty. As the case study playlists show, music variety narrowed after the playlist tool's introduction; Figure 9 confirms that music segment lengths became increasingly consistent, indicating a more controlled, computer-calculated playlist structure.

5: Discussion and comparison with qualitative study

This research examined general developments in P3 morning programming (6 a.m. to 12 p.m., 1989-2019) by measuring music and speech proportions and their respective segment lengths. Figures 6 through 9 reveal visually striking changes in the years 1992, 2001, 2005, and 2016.

Several institutional and media-political shifts help explain these changes. The official end of DR's monopoly arrived in 1983 when local radio stations were first permitted, but the monopoly for nationwide channels was not fully broken until 2003. Thus, P3 faced growing competition over three decades, increased further by streaming services such as Spotify and iTunes from the late 2000s.

To counter competition from local and commercial music stations, P3's profile was strengthened through the 1992 restructuring of DR's four radio channels: P1, P2, P3, and Danmarkskanalen (renamed P4 in 1997). Along with a more market-oriented approach, a controversial playlist tool for music control was introduced and gradually implemented alongside rotation systems during the 1990s. This meant that music presenters no longer chose the tracks themselves. This systematization and standardization of musical track lengths likely explains the increase in average music length in 1992, supported by the growth in music segments of around 200 seconds after 1995 visible in Figure 9.

In Figure 6, significant changes around 1992, 2001, and 2005 align with the case study's finding of less speech in the early 1990s before the restructuring. A growth in speech peaked around 2011 before declining towards 2016. Notably, the large-scale study shows an upward trend in speech from 2016 to 2019, coinciding with decreased musical content.

The rise in speech from 2005 to 2011 in Figure 6 can be linked to more hosts in the studio (see Figure 10). More hosts produce more talk. Have (2018, p.132f) documented many different hosts in the early years, but typically one co-present (1989-2000). A stable group of three to four hosts appeared (2005-2015), with a brief interim of one or two hosts (2001-2004).

The 2005 changes might also be due to a data format shift. That year, the Danish State Media Collection moved from digitizing incoming provider material to receiving born-digital files, as previously described. While positive, this shift led to analysis focusing on individual programs instead of the full morning segment, resulting in files dominated by either music or speech and thus increased uncertainty. To normalize across the two corpora, each file was divided into segments of roughly 1,000 seconds, reducing outliers. Nonetheless, a noticeable difference persists. As Figures 8 and 9 show, hardly any speech or music segments exceed 1,000 seconds, but increased uncertainty after 2005 must still be accounted for.

Figure 7 shows average music length rising from 1992 and both music and speech lengths declining from 2011. This suggests shorter segments of both, possibly explained by hosts overlapping with music, causing more frequent switching. The 2011 shift could result from increased use of jingles and segments like quizzes or DR's self-promotion—though that does not align with Figure 4, which shows a fairly stable "Other" category between 2011 and 2016. The increase in speech length from 1992 to 1999 likely stems from DR's 1992 reorganization, which introduced more entertaining, chatty news communication as part of Go'd Morgen P3, occupying half of the morning programs over the study period. After 1992, P3 gradually transformed from a purer music channel into a more entertaining, communicative one.

Figures 8 and 9 again highlight the first peak of very short segments (10-20 seconds) due to overlapping host talk, and the second peak around 200 seconds—especially for music. This confirms standardized musical track lengths after the 1992 restructuring, an instance of what Have (2018) calls radiotization of music. In other words, radio's logic now influences musicians' creative choices. To be played on air, tracks must run close to 3 minutes 20 seconds to fit the flowing musical clock structure (see e.g. Hendy 2000, 94ff; Michelle et al. 2018a, p.192f; Russo 2013). Segments slightly above 3 minutes have proven optimal for holding listeners—a balance of recognizability and novelty. Playlist in the case study show narrower music variation after the playlist tool's adoption; Figure 9 reinforces this by revealing increasingly stable music segment lengths.

Besides specific findings on Danish radio channel P3, this article also explores methodology for combining deep and distant listening. By moving flexibly between the existing case study and a larger-scale analysis—an approach that might be called meso-scale listening—the analysis has become not only more robust but also fills gaps inherent in each method. Risks of cherry-picking in the qualitative case study were countered, and Figure 6 added nuance by revealing considerable year-to-year variation in speech, including an upward trend from 2016 onward—a period not covered by the case study. This raises new questions, such as why speech increased after 2016. One explanation might be increasing competition from commercial digital music radio stations and streaming services. In this landscape, DR and P3 emphasize their core public service competency: professional journalistic content delivered by well-known young personalities in an entertaining style. This strategy distances P3 even further from its origins as a music channel.

Many visible changes in Figures 6-9 can be linked to institutional shifts at DR, but the notable change in 2005 defies such explanation. This prompted questioning not of the classifier but of the data, whose format changed that year. This insight is valuable for future large-scale studies using LARM.fm. Working with the archive as reverse-engineering humanists brought critical attention to significant oscillations not always tied to actual content changes but rather to the data formats providing that content.

A principal aim of this overall study was to analyze whether the content has shifted from music-dominated to more spoken word. The close-listening approach enabled examination of how music was presented by hosts, confirming a shift from hosts filling gaps between tracks by talking about music in the early 1990s to music filling gaps between a group of hosts who rarely reference it. However, the large-scale study corrected this narrative by showing that, overall, speech proportion has not steadily increased; rather, it has fluctuated over thirty years, as seen in Figure 6. The diachronic change in music versus speech is less drastic than initially implied. Single case studies—even deliberately sampled—can lead to questionable conclusions when not examined against the full programming landscape.

Combining the two approaches generates more valid answers, clearly illustrating each method's strengths and weaknesses.

6: Conclusion: Combining close to distant listening

This study compared a case study of five Danish music radio programs (1990-2016) with a large-scale analysis of P3's entire morning programming (6 a.m. to 12 p.m., 1989-2019). Both drew on the LARM.fm digital archive, which offers new pathways for radio and media studies by affording both deep and distant listening. This is the first large-scale analysis of such a volume of Danish radio material. Following Papakostas and Giannakopoulos, we applied a CNN-based image classification on spectrograms, achieving state-of-the-art performance, and comparative results with an SVM classifier by Giannakopoulos (2015). As Table 2 and Figure 5 demonstrate, these classifier tools delivered high accuracy that generalized well to Danish. This finding advances automated speech recognition in digital humanities: for instance, showing how findings from small datasets can be scaled up and how speech-music discrimination models transfer across languages (here Danish and English). The model may also be adaptable to tasks such as gender or regional accent detection.

The findings confirm that political and institutional shifts in Danish music radio leave a signature on program content. For example, standardized formats and segments emerge after the 1992 restructuring. Still, the most salient spike in Figure 6 in 2005 is more attributable to a format change from digitized to born-digital files than to actual content change. This study thus demonstrates the necessity of reflecting on data constitution and change when undertaking large-scale analysis.

The synergistic pairing of close listening and distant listening has provided a more richly detailed picture of how morning music radio programming on P3 evolved from 1989 to 2019. This methodological combination not only yields more valid answers regarding shifts in the distribution of music and talk over the three decades, but also exposes distinctive strengths and limitations of both the qualitative case study and the large-scale computational analysis. Ultimately, we hope this work inspires Digital Humanities researchers to incorporate more audio content analysis into Media Studies inquiry.