Foreign Accent and Forensic Speaker Identification in Voice Lineups: The Influence of Acoustic Features Based on Prosody
September 27th, 2024
This research aimed to make a comparison between L1-L2-English and L1-L2 Portuguese to check how much the effect of a foreign accent accounts for both metrics and prosodic-acoustic parameters, as well as for the choice of the target voice in a voice lineup.
Our research examines how a foreign accent influenced speaker identification. We focus on prosodic features based on the fundamental frequency, which is the voice pitch, duration, and voice quality. Our goal is to understand how these features influence listeners'judgments in voice lineups.
There is an increasing focus and research into the performance of automatic speaker recognition, which applies automation to the workflow of forensic speaker comparison. However, the information is like a black box for forensic scientists to report to the police, judges, and jurors. Automatic speaker recognition systems based on classic techniques such as GMM, UBM models, and live actors.
There is also neural research based on artificial intelligence. We propose an automated flow preserving the linguistic information missed by the automatic speech recognition systems. Our protocol uses a combined auditory and acoustic approach for forensic speech comparison while establishing countries where the science was developed, but using automated tools for extracting a wide range of acoustic features as well as for running acoustic similarity procedures.
To begin, write the linguistic transcription for each audio file in a TXT file format. Tag the pair of TXT and WAV files with the same name. Create a folder for each L1, L2 language.
Ensure that all file pairs of the same language are in the same folder. Access the web interface of Munich Automatic Segmentation forced aligner, drag and drop each pair of WAV and and TXT files from the folder to the dashed rectangle in files. Click the upload button to upload the files to the aligner.
In the service options menu, for L1 L2 English data, select grapheme to phoneme to mouse to phone to syllable for pipeline name and English-US for language. Keep the default options for output format and keep everything. Check the run option box to accept the terms of usage.
Click the run web service button to run the uploaded files in the aligner. After the files are processed, click the download as zip file button to download the text grid files. Extract the text grid files for later realignment in phonetic analysis software.
Access and download the script for PRAAT VVUnitAligner. Ensure that all file pairs of the same language and the VVUnitAligner script are in the same folder. Open the phonetic analysis software.
From the object window, click Praat and open Praat script to load the script. Click the run button, then select the language as English-US. Now, from the chunk segmentation button, select automatic.
Check the save text grid files option to automatically save the newly generated text grid files. Click the okay and run buttons for realignment of the phonetic units. From the given site, download the speech rhythm extractor script for automatic extraction of prosodic acoustic features.
Create a new folder and add the speech rhythm extractor script along with all audio text grid files of all languages. Open the phonetic analysis software. From the object window, click Praat and open Praat script to load the script.
Then click the run button once. Check the voice quality parameters option to save the output file VQ for voice quality. Now check the linguistic target option to choose the language.
Then check the unit option to choose the F0 features in semitones. Set the values for the F0 threshold, including minimum and maximum thresholds. Click okay, followed by run for the automatic extraction of acoustic features.
To perform generalized additive models, non-parametric statistical analysis, type the indicated command and upload the spreadsheet containing the extracted acoustic features into the R environment. Finally, press enter to execute. Speech rate decreased more rapidly for L1 L2 English compared to L1 L2 BP, which had less steep slopes due to higher syllable duration and lower variability.
Local shimmer remained relatively stable for Brazilian speakers, L1 BP and L2 English, despite increasing syllable duration variability. The pause rate was higher for L2 BP speakers, with longer pauses compared to L1 English, L1 BP, and L2 English speakers. Articulation rate was similarly affected as the speech rate with lower rates associated with higher cognitive linguistic load and syllable variation.
The standard deviation of syllable duration decreased as speech rate increased across all language levels. Varco of syllables decreased for L1 BP and L2 BP with increasing F0 variability and speech rate while it increased for L1 English and L2 English. The standard deviation of consonants showed lower variability in L1 BP as speech rate or pause duration increased compared to L1 English.
The standard deviation for vowels and consonants followed a fall rise pattern for L1 BP and L2 BP, with increasing prosodic features, while it decreased and then attenuated for L1 English and L2 English. After preparing four voice lineups each for English and BP, get the audio files from the selected speakers and arrange them into language specific folders. Randomly select six voice chunks in L1 English or L1 BP.Then choose one voice chunk in L2 English or L2 BP from one of the six voice chunks.
Access and download the script for Praat Create Lineup. Before running the script, ensure that the L2 reference voice, L1 foils, and L1 target voice are placed in the same folder. Open the phonetic analysis software.
From the object window, click Praat and open Praat script to load the script. Then click run to execute the create lineup script. In the R environment, to perform the Kruskal-Wallace test, type the indicated command.
Then upload the spreadsheet containing the scores of the listeners'judgments and press enter. Then for post-hoc Dunn's test, type the following command and press enter. Access and download the Python script, Acoustic Similarity Cosine Euclidean.
Ensure that the downloaded script is saved in the same folder as the voice lineup dataset. Click the open file button to call the script, then click run, and run without debugging buttons to execute the script. Finally, perform voice similarity tests based on acoustic features.
In BP voice lineup one, foil voice three was judged as the target voice, with no significant difference between foil three and target voice four. In BP voice lineup two, no significant difference was found between target voice three and foil four. Both cosine similarity and Euclidean distance showed a strong correlation between foil three and the target voice in BP lineup one.
In BP lineup two, both similarity metrics correlated strongly between foil four and the target.
Overview
This research examines how a foreign accent influences speaker identification, focusing on prosodic features such as voice pitch, duration, and quality. The study aims to understand how these features affect listeners' judgments in voice lineups.
Key Study Components
Area of Science
- Neuroscience
- Psychology
- Forensic Science
Background
- Research on automatic speaker recognition is increasing.
- Prosodic features play a crucial role in speaker identification.
- Forensic scientists often face challenges in interpreting automatic recognition systems.
- Understanding listener judgments can improve forensic practices.
Purpose of Study
- To compare the effects of foreign accents on speaker identification.
- To analyze prosodic-acoustic parameters in voice lineups.
- To enhance the understanding of listener judgments in forensic contexts.
Methods Used
- Analysis of fundamental frequency and voice pitch.
- Examination of duration and voice quality metrics.
- Comparative study between L1-L2-English and L1-L2-Portuguese.
- Evaluation of listener responses in voice lineups.
Main Results
- Foreign accents significantly affect speaker identification.
- Prosodic features influence listener judgments.
- Differences were noted between English and Portuguese speakers.
- Findings contribute to the understanding of forensic speaker comparison.
Conclusions
- Accent impacts listener perception in voice lineups.
- Prosodic features are critical in speaker recognition tasks.
- Further research is needed to refine forensic methodologies.
Frequently Asked Questions
Executive Industry Relevance
Quantitative analysis of prosodic-acoustic features in foreign-accented speech enables objective assessment of speaker identity, supporting forensic and security-focused R&D pipelines. Integrating statistical modeling and acoustic similarity metrics enhances predictive confidence in voice-based identification, reducing ambiguity in high-stakes decision points. These capabilities are critical for enterprise workflows where robust, reproducible speaker comparison informs risk-adjusted portfolio decisions.
Strategic Applications in Biopharma R&D
Early Discovery & Target Validation
- Enables rigorous hypothesis testing of prosodic feature impact on speaker differentiation.
- Supports mechanistic de-risking by quantifying variability in f0, duration, and voice quality.
- Facilitates functional validation of acoustic markers relevant to identity verification.
Screening & Assay Development
- Standardizes extraction of acoustic features for reproducible downstream analysis.
- Prepares validated datasets for scalable screening of voice samples in forensic contexts.
- Enables quantitative comparison of lineup voices, supporting reliable evaluation protocols.
Translational & Preclinical Research
- Aligns acoustic feature analysis with translational needs in forensic and security applications.
- Ensures continuity from discovery of prosodic markers to preclinical validation of identification systems.
- Supports risk-adjusted advancement of voice-based biomarker strategies.
Pipeline & Workflow Integration
This protocol integrates from early discovery of acoustic markers through statistical validation and screening, supporting workflows from hypothesis testing to preclinical system evaluation.
- Discovery Biology: Quantifies prosodic-acoustic features to clarify mechanisms underlying speaker identification.
- Screening: Provides standardized, reproducible feature extraction for lineup evaluation.
- Analytics: Delivers statistical outputs (GAMs, Kruskal-Wallis, Dunn's test) and similarity metrics for robust condition comparison.
- Translational Research: Bridges discovery of acoustic markers to operational forensic applications.
- Enterprise Reuse: Establishes a reusable analytical pipeline for voice-based identification studies.
Operational & Enterprise Impact
- Scientific Value: Increases predictive confidence and reduces mechanistic ambiguity in speaker identification.
- Operational Value: Delivers standardized, scalable, and reproducible acoustic analysis workflows.
- Strategic Value: Informs go/no-go decisions and enhances capital efficiency in forensic R&D portfolios.
- Portfolio Impact: Supports risk-adjusted prioritization of voice-based identification technologies.
Implementation Considerations
- Requires expertise in acoustic analysis, statistical modeling, and phonetic software.
- Depends on access to specialized instrumentation and analytical infrastructure (e.g., PRAAT, R, Python scripts).
- Demands cross-team standardization of data formats and feature extraction protocols.
- Must be adapted for language-specific and accent-specific model systems.
- Practical limitations include variability in listener perception and language group differences.
