179566 research outputs found
Sort by
“The hunger games”:Politics and drought in post-colonial Zimbabwe, 1980-1992
Get PDFThis paper explores the politics of drought and hunger in the immediate post-colonial period in Zimbabwe, 1980–1992. This period was characterised by the state's efforts to consolidate power in the context of the African citizenry's anticipated reform of the colonial system and ensuing white supremacist state. The new African-led state's power consolidation efforts manifested in the Gukurahundi massacre between 1982 and 1987, in which drought became a tool to enforce hunger in the Midlands and Matabeleland provinces. There followed a continuation of the politics of patronage and poor administration, which came to characterise the state's drought relief programmes. We explore this political context and deploy a socio-environmental approach to explain the history of drought in the first decade of post-independence Zimbabwe. We begin our analysis in 1980 at the start of political independence, where the political structure of drought response changed, and end in 1992 with the worst drought of the century. Our primary sources are oral history interviews, newspapers, and parliamentary debates. We demonstrate that the post-colonial droughts between 1980 and 1992 were characterised not only by failure at government level but were a conscious manipulation in line with the ruling party's neopatrimonialism stance. Where loyalty could not be bought, it had to be coerced. So the state-forced starvation on drought victims was central to its strong-arm programme of control. Indeed, the failure of the rains could be an excuse when the state needed to claim an alibi for its own failures. In essence, we contend that politics eclipsed policy in fighting drought
For God and Empire:Arabic at the University of Edinburgh: its development, character and mission
Get PDFArabic was first taught at the University of Edinburgh in 1751, but it was not officially introduced as a regular subject until 1859 when it became an extra subject of study for the Senior Hebrew class. Over the following decades, it was taught by professors of Hebrew (and Semitic languages) until a dedicated lecturer in Arabic was appointed in 1911. Thereafter, it continued to be taught within the framework of Semitic languages until the early 1950s when a Department of Arabic was established, along with Turkish and Persian departments, in response to changing post-war needs. This chapter explores the development of the teaching of Arabic at the University of Edinburgh from the mid-nineteenth century until the early 1950s, looking at its practitioners, curriculum, degree, and student body. Originating from a milieu of religious studies and continental philology that also served staffing needs of imperial administration, it was taught largely by churchmen for much of the period and it was not until the early 1950s that the first native speaker of the language was employed on staff. It further examines the changing character of Arabic teaching at Edinburgh, the programme of study, its approach to spoken Arabic, and other changes within the curriculum that saw it begin to move slowly away from a traditional orientalist construct towards a more area studies model
Segmentation-Variant Codebooks for preservation of paralinguistic and prosodic information
No full textDiscrete speech units (DSUs) from self-supervised models like HuBERT are widely adopted in speech modeling, compression, and generation. While effective for lexical content-heavy tasks like speech recognition, understanding, and Speech Emotion Recognition (SER), DSUs often lack crucial prosodic and paralinguistic information for expressive speech generation (Ngyuen et al. 2023). Increasing the codebook size can mitigate this information loss but leads to inefficiently higher bitrates. Inspired by SUPERB's utterance-level pooling of continuous SSL features for SER, we propose Segmentation-Variant Codebooks (SVCs): a method that quantizes speech at multiple linguistic levels (frame, phone, word, and utterance) using separate codebooks, producing parallel streams of segment-specific discrete representations. Phone and word boundaries are extracted using the Montreal Forced Aligner, then final-layer HuBERT features are pooled over these boundaries, as shown in Figure 1. Each pooled set is quantized using a KMeans++ codebook specific to its segmentation level, resulting in a variable bitrate reflecting speech rate and corpus-specific characteristics. Table 1 shows average bitrates across datasets, comparing SVCs with fixed-frame DSU baselines. For classification, we use a single linear layer followed by softmax (emotion) or sigmoid activation (prominence). Frame-level probing on IEMOCAP (emotion) and Naver Prosody Control (prominence) Table 2 shows that SVC-derived DSUs retain more salient, linearly separable paralinguistic and prosodic information compared to fixed-frame DSUs. For resynthesis, we follow (Ngyuen et al. 2023) by training a HiFi-GAN with speaker conditioning. We evaluate synthesis quality using UTMOS, intelligibility via WER (using Whisper-large for transcription), and style preservation with the pretrained classifier from (Ngyuen et al. 2023). As shown in Table 3, SVCs outperform fixed-frame DSUs at comparable or lower bitrates, significantly reducing the gap with continuous representations. SVC-based resynthesis yields minor improvements in intelligibility and perceptual quality, while yielding larger improvements in expressive variation compared to frame-level baselines
Access Repository Dashboard
Do you manage Edinburgh Research Explorer? Access insider analytics, issue reports and manage access to outputs from your repository in the CORE Repository Dashboard!