Innovative microphone transmitter reveals differences in acoustic structure between broadcast and whisper songs of Myadestes obscurus (ʻŌmaʻo)

Low-amplitude ‘whisper songs’ are a taxonomically broad phenomenon in birds that could play an important role in the suite of behaviors birds use to communicate. Due to its cryptic nature, there are inherent difficulties in capturing high-quality whisper song recordings without interrupting natural behaviors. Thus, whisper song acoustic structure is poorly understood and its potential function remains the subject of debate. Here, we present one of the first quantitative assessments of the acoustic structure of whisper song in birds. Using an innovative microphone transmitter we collected high quality recordings of broadcast and whisper songs from the Myadestes obscurus (ʻŌmaʻo), a thrush species endemic to the Island of Hawai‘i. The transmitter was attached to the birds and broadcasted radio-signals of all vocalizations produced by the individual to distances over 100 m away which minimized disruption of the birds’ normal behavior while recording. We demonstrate that ʻŌmaʻo whisper songs are a distinct class of vocalization that differ from broadcast songs in acoustic characteristics beyond amplitude, such as song length, frequency, and length of silent intervals between notes. These findings, in conjunction with habitat-associated variation in the rate at which ʻŌmaʻo emit these vocalization classes, indicate broadcast and whisper songs likely serve separate functions. This work provides evidence supporting the ‘acoustic adaptation hypothesis’ which posits that densely vegetated habitats promote the evolution of songs with specific acoustic features that maintain signal integrity as the sound propagates through the environment.

Methods
We equipped ʻŌmaʻo with a device that consisted of a very high frequency (VHF) radio pulse transmitter and a miniature condenser microphone that broadcasted all sounds emitted by the target bird (JDJC Corporation, Fisher, IL, USA). Radio signals from the microphone transmitter were received by a hand-held 3-element yagi antenna (JDJC Corporation, Fisher, IL, USA) connected to a wideband receiver (AR8200; AOR Ltd., Tokyo, Japan) in wideband amplitude modulation (WAM) mode. The signals were recorded by a Marantz PMD661 professional field recorder (Marantz America, LLC., Mahwah, NJ, USA), using a 2-second pre-record setting, as 24-bit WAV files at a 44.1 kHz sampling rate. The distance at which the microphone transmitter could be detected varied by terrain and weather, and birds recorded in our study were typically at a distance of < 80 m. Each bird was tracked 3–5 days a week, excluding rainy days, for 2–4 hours between 0800–1600, which encompasses the peak hours for ʻŌmaʻo vocalizations. 
Recordings were processed using the bioacoustic software Luscinia (v2.16.10.29.0). Songs were delineated by > 5 s of silence between vocalization bouts and classified as either broadcast or whisper songs based on field observations, with verification through spectrogram visualization. Songs were then segmented into syllables, delineated by > 0.2 s of silence between notes in a song, and syllables were further segmented into notes, delineated by any break in continuous sound production. In Luscinia, frame length was set to 5 ms, time step was set to 1.1 ms, maximum frequency was set to 8 kHz, and the Hamming windowing function was selected before each note contour (vectors of parameter measurements with one entry for each time step in the spectrogram) was manually traced with a brush size of 5 pixels. 
To determine the number of unique syllables associated with broadcast and whisper songs for each bird, we used Luscinia’s dynamic time warping (DTW) algorithm. The DTW algorithm compares the structure of each syllable with every other syllable based on the Euclidean distance between acoustic features, producing a matrix of syllable dissimilarities. To achieve optimal syllable groupings, DTW acoustic parameters were weighted differently for broadcast and whisper songs. We clustered syllables using unweighted pair group method with arithmetic mean (UPGMA) hierarchical clusterings of the dissimilarity scores from the DTW analysis, assigning syllables to groups based on the resulting dendrogram. We calculated the global silhouette index, a measure of clustering tendency, at each depth within the dendrogram and identified natural clusters among the syllables by searching for the highest peak in the index. The syllable groupings were then visually and aurally validated and reassigned as necessary by an avian acoustics specialist. Although different syllables were incorrectly assigned to the same group < 1% of the time, similar syllables were frequently separated into different groups and needed manual grouping. Syllables were classified into separate groups if they had two or more differences in the number of notes or note shape.