Phonetic & speech analysis
Segment, annotate and measure speech recordings (formants, pitch, durations).
20 project-authored proposals · taxonomy v1.1 · 0 new benchmark runs. Verifiers and partial-credit rules below are designs, not executed results.
- Typical applications
- Praat · ELAN · Audacity
- Class GUI-need label
HY· Hybrid
Task proposals · 20
Expand to inspect the complete original wording and proposed checks. Original proposal text remains English in both interfaces.
R08.10.01.04-T01 Segment 10 CVC words into word and phone tiers in Praat
Original instruction
Open input/vowels_f01.wav (female speaker, 10 CVC words in carrier phrases) in Praat. Create output/vowels_f01.TextGrid with interval tiers 'word' (target words only) and 'phone' (C, V, C of each target, SAMPA labels from input/wordlist.txt), placing boundaries from the spectrogram and waveform.
- Platforms
mac· macOSwin· Windows- Applications
- Praat
- Interfaces
desk· Desktop GUI- GUI need · task setting
G2· GUI-essential- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
av· Audio/video over timedrag· Drag & dropvis· Visual judgment- Work product
WP10· Annotations- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Checker parses the TextGrid: labels exact; >=90% of phone boundaries within 20 ms of gold (two-annotator consensus), all word boundaries within 30 ms. Forced-aligner output is detected by its characteristic 10 ms frame-quantized boundaries and fails.
Proposed partial credit
20% tiers and labels exact; 30% word boundaries; 50% phone boundaries within 20 ms (scaled, full credit at 0.9 of boundaries)
R08.10.01.04-T02 Measure F1 and F2 at vowel midpoints from an existing TextGrid
Original instruction
input/speaker_m03.wav comes with input/speaker_m03.TextGrid (40 labelled vowels on tier 'phone'). In Praat measure F1 and F2 at each vowel's temporal midpoint using Burg with max formant 5000 Hz, 5 formants, 25 ms window, and write output/formants.csv (label, start, end, F1, F2).
- Platforms
mac· macOSwin· Windows- Applications
- Praat
- Interfaces
desk· Desktop GUI- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
CL· CLI-sufficient- Interaction primitives
code· Code & commandsclick· Point & click- Work product
WP02· Records/table- Conditions
- None declared
- Proposed difficulty
- 2 / 5
Proposed verifier
Checker recomputes with parselmouth using the stated settings: all 40 rows present, F1 within 30 Hz and F2 within 60 Hz of reference; intervals must match the TextGrid times exactly.
Proposed partial credit
20% 40 rows with correct times; 40% F1 within tolerance; 40% F2 within tolerance
R08.10.01.04-T03 Correct octave errors in a creaky speaker's pitch track
Original instruction
input/creaky_f12.wav has creaky phrase endings that produce pitch halving and doubling. In Praat choose suitable pitch floor and ceiling, then open a Manipulation object and remove or correct the 6+ octave errors in the PitchTier by hand. Save output/corrected.PitchTier.
- Platforms
mac· macOSwin· Windows- Applications
- Praat
- Interfaces
desk· Desktop GUI- GUI need · task setting
G2· GUI-essential- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
drag· Drag & dropav· Audio/video over timevis· Visual judgment- Work product
WP10· Annotations- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
Checker samples the PitchTier every 10 ms in voiced frames of the gold track (EGG-derived): RMSE<=5 Hz and no point deviating by more than 30% (octave jump). Uncorrected default analysis gives RMSE ~38 Hz. Praat's New/Open Praat script commands are hidden via its Buttons preferences and the VM has no shell for praat --run, so points must be edited by hand in the Manipulation window.
Proposed partial credit
20% PitchTier saved covering the file; 40% no octave jumps remaining; 40% RMSE<=5 Hz
R08.10.01.04-T04 Annotate speaker turns and overlaps in a conversation video in ELAN
Original instruction
Open input/dinner_conv.mp4 (3 min, two speakers A and B) in ELAN with template input/turns.etf. Create annotations on tiers 'A_turn' and 'B_turn' for every turn (including backchannels) and on tier 'overlap' for each stretch of simultaneous speech. Save output/dinner_conv.eaf.
- Platforms
mac· macOSwin· Windows- Applications
- ELAN
- Interfaces
desk· Desktop GUI- GUI need · task setting
G2· GUI-essential- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
av· Audio/video over timedrag· Drag & dropkey· Shortcuts- Work product
WP10· Annotations- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
Checker parses the EAF: turn boundary F1>=0.80 at 200 ms tolerance per tier, overlap temporal IoU>=0.6; media link must point to the video. Tier names and types from the template unchanged. ELAN has no scripting API and the VM has no shell or text editor; .eaf writes by any process other than ELAN are caught by file-access auditing and fail.
Proposed partial credit
15% EAF with template tiers and media link; 50% turn boundary F1; 35% overlap IoU
R08.10.01.04-T05 Set up an ELAN morpheme-gloss tier hierarchy and enter glosses
Original instruction
Open input/narrative.eaf (20 utterances on tier 'tx'). Add a child tier 'mb' (Symbolic Subdivision) and 'ge' (Symbolic Association under 'mb') using linguistic types from input/igt.etf, then enter morpheme breaks and glosses for all utterances from input/glosses.xlsx. Save output/narrative.eaf.
- Platforms
mac· macOSwin· Windows- Applications
- ELAN
- Interfaces
desk· Desktop GUI- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
CL· CLI-sufficient- Interaction primitives
click· Point & clicktype· Text & forms- Work product
WP10· Annotations- Conditions
- None declared
- Proposed difficulty
- 2 / 5
Proposed verifier
Checker validates EAF tier parents and constraint stereotypes exactly; morpheme and gloss sequences per utterance equal glosses.xlsx; original 'tx' annotations unchanged.
Proposed partial credit
35% tier hierarchy and constraints exact; 50% glosses correct (proportional); 15% tx unchanged
R08.10.01.04-T06 Denoise, normalize and label stimulus tokens in Audacity
Original instruction
input/stimuli_raw.wav has 12 spoken tokens with fan noise. In Audacity, take a noise profile from 0-1.5 s, apply Noise Reduction (12 dB), normalize peak to -1.0 dBFS, and add a label track marking each token's start and end with labels from input/tokens.txt. Export output/stimuli_clean.wav and output/labels.txt.
- Platforms
mac· macOSwin· Windows- Applications
- Audacity
- Interfaces
desk· Desktop GUI- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
CL· CLI-sufficient- Interaction primitives
drag· Drag & dropav· Audio/video over timeclick· Point & click- Work product
WP05· Time-based media- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Peak of output WAV -1.0+-0.1 dBFS; SNR (token vs silence RMS) improved >=6 dB vs input; labels exact text in order and boundaries within 30 ms of gold onset/offset; length and sample rate unchanged.
Proposed partial credit
25% noise reduction SNR gain; 15% peak level; 60% labels and boundaries (proportional)
R08.10.01.04-T07 Measure voice onset time for 30 stop tokens in Praat
Original instruction
input/vot_session.wav contains 30 word-initial stops (p t k b d g). In Praat create a point tier 'burst' and a point tier 'voicing' marking release burst and onset of periodic voicing (prevoicing onset for voiced stops), then save output/vot.TextGrid and output/vot.csv (token, VOT_ms, negative for prevoicing).
- Platforms
mac· macOSwin· Windows- Applications
- Praat
- Interfaces
desk· Desktop GUI- GUI need · task setting
G2· GUI-essential- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
av· Audio/video over timevis· Visual judgmentdrag· Drag & drop- Work product
WP02· Records/table- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
Each VOT within 5 ms of gold (phonetician consensus); sign correct for prevoiced tokens; CSV values must equal differences of TextGrid points within 1 ms. Praat's New/Open Praat script commands are hidden via its Buttons preferences and the VM has no shell for praat --run; output writes by other processes are caught by file auditing.
Proposed partial credit
20% 60 points placed; 20% sign correct for all tokens; 60% VOT within 5 ms (proportional)
R08.10.01.04-T08 Fix misaligned phone boundaries in a forced-alignment TextGrid
Original instruction
input/reading_s07.TextGrid came from a forced aligner for input/reading_s07.wav; 15 phone boundaries are off by 30-120 ms (fricative-vowel and nasal transitions). In Praat inspect the spectrogram and move the wrong boundaries; do not change labels or the word tier. Save output/reading_s07_fixed.TextGrid.
- Platforms
mac· macOSwin· Windows- Applications
- Praat
- Interfaces
desk· Desktop GUI- GUI need · task setting
G2· GUI-essential- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
drag· Drag & dropav· Audio/video over timevis· Visual judgment- Work product
WP10· Annotations- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
Mean absolute error of the 15 planted boundaries <=15 ms vs gold; other boundaries moved by <5 ms; labels and word tier identical. Re-running an aligner is detected by many unplanted boundaries shifting.
Proposed partial credit
60% planted boundaries corrected (proportional); 25% untouched boundaries stable; 15% labels/word tier identical
R08.10.01.04-T09 Control pair of T02: batch formant extraction with a Praat script
Original instruction
Control pair of T02. input/corpus/ holds 200 WAV+TextGrid pairs from 10 speakers. Write output/formants.praat that runs headless (praat --run) and writes output/all_formants.csv with F1 and F2 at vowel midpoints, using max formant 5000 Hz for male and 5500 Hz for female speakers (sex in input/speakers.csv).
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- Praat
- Interfaces
term· Terminal- GUI need · task setting
CL· CLI-sufficient- GUI need · ordinary access
CL· CLI-sufficient- Interaction primitives
code· Code & commands- Work product
WP02· Records/table- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Checker runs the script on the corpus and a held-out copy with 20 new files; row counts exact; F1/F2 within 30/60 Hz of parselmouth reference with the sex-specific ceiling.
Proposed partial credit
20% script runs headless; 20% row counts; 30% sex-specific settings applied; 30% values within tolerance
R08.10.01.04-T10 Plot a vowel space with category ellipses and compute its area
Original instruction
Using input/vowels_5spk.csv (F1, F2 for 5 vowel categories), make a Praat Picture of the F2-F1 vowel space (both axes reversed, Hz, 1-SD ellipses per vowel, IPA labels at means) and save output/vowel_space.pdf; also save output/space.json with category means and the convex-hull area of the 5 means in Hz^2.
- Platforms
mac· macOSwin· Windows- Applications
- Praat
- Interfaces
desk· Desktop GUI- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
CL· CLI-sufficient- Interaction primitives
code· Code & commandsvis· Visual judgment- Work product
WP03· Visual layout- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Means within 1 Hz and hull area within 1% of reference. PDF parsed: axes reversed (tick order), 5 ellipse paths and 5 labels present; VLM rubric only for label placement (10%).
Proposed partial credit
30% category means; 30% hull area; 30% PDF axes/ellipses/labels; 10% VLM placement rubric
R08.10.01.04-T11 Compute spectral moments for sibilant fricatives
Original instruction
input/sibilants.wav with input/sibilants.TextGrid marks 24 /s/ and /S/ tokens. For each token compute the spectral center of gravity, SD, skewness and kurtosis over the middle 40 ms (Hamming window, 500 Hz high-pass) in Praat and write output/moments.csv; then report the mean CoG difference /s/-/S/ in output/answer.txt.
- Platforms
mac· macOSwin· Windows- Applications
- Praat
- Interfaces
desk· Desktop GUI- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
CL· CLI-sufficient- Interaction primitives
code· Code & commandsclick· Point & click- Work product
WP02· Records/table- Conditions
- None declared
- Proposed difficulty
- 2 / 5
Proposed verifier
Checker recomputes with the stated settings: CoG within 50 Hz, other moments within 5% per token; answer within 50 Hz.
Proposed partial credit
20% 24 tokens with correct windows; 50% moments within tolerance; 30% answer
R08.10.01.04-T12 Annotate gesture phases frame-accurately in ELAN
Original instruction
input/lecture_gestures.mp4 (25 fps) shows a speaker making 10 co-speech gestures. In ELAN create tier 'phase' with annotations Prep, Stroke, Hold, Retraction for each gesture, stepping frame by frame to set boundaries, and link each stroke to the word it co-occurs with on tier 'affiliate'. Save output/gestures.eaf.
- Platforms
mac· macOSwin· Windows- Applications
- ELAN
- Interfaces
desk· Desktop GUI- GUI need · task setting
G2· GUI-essential- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
av· Audio/video over timekey· Shortcutsnav· Pan/zoom/scroll- Work product
WP10· Annotations- Conditions
- None declared
- Proposed difficulty
- 5 / 5
Proposed verifier
Phase labels match gold sequence per gesture; boundaries within 2 frames (80 ms) for >=80% of boundaries; affiliate words exact for >=8 of 10 strokes. ELAN has no scripting API and the VM has no shell or text editor; .eaf writes by any process other than ELAN are caught by file-access auditing and fail.
Proposed partial credit
20% 10 gestures segmented; 20% phase label sequences; 40% boundary precision; 20% affiliates
R08.10.01.04-T13 Label ToBI pitch accents and boundary tones for 8 utterances
Original instruction
Open input/tobi_set.wav with input/tobi_set.TextGrid (word tier only) in Praat. Add point tier 'tones' and label pitch accents (H*, L*, L+H*, L*+H, H+!H*, !H*) and phrase/boundary tones (L-, H-, L%, H%) following MAE_ToBI conventions, plus a 'breaks' tier with break indices 0-4.
- Platforms
mac· macOSwin· Windows- Applications
- Praat
- Interfaces
desk· Desktop GUI- GUI need · task setting
G2· GUI-essential- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
av· Audio/video over timevis· Visual judgmenttype· Text & forms- Work product
WP10· Annotations- Conditions
tutorial· Follow docs/tutorial- Proposed difficulty
- 5 / 5
Proposed verifier
Checker aligns tone points to words (within 50 ms): pitch-accent presence F1>=0.80, accent-type agreement kappa>=0.6, boundary-tone accuracy>=0.8 and break-index agreement within +-1 for >=90% of words vs expert gold. Praat's New/Open Praat script commands are hidden via its Buttons preferences and the VM has no shell for praat --run; TextGrid writes by other processes are caught by file auditing.
Proposed partial credit
15% tiers present; 30% accent presence F1; 25% accent types; 30% boundary tones and break indices (2:1)
R08.10.01.04-T14 Split a 45-minute session into sentence files with TextGrids
Original instruction
input/session.wav is a 45-min read-speech session of the 120 sentences in input/prompts.txt, with repetitions and false starts. Use Audacity and/or Praat to cut one file per prompt (keep the last fluent reading, 150 ms padding), name them s001.wav..s120.wav, and create a matching TextGrid with a single interval containing the prompt text.
- Platforms
mac· macOSwin· Windows- Applications
- Audacity · Praat
- Interfaces
desk· Desktop GUIterm· Terminal- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
CL· CLI-sufficient- Interaction primitives
av· Audio/video over timedrag· Drag & dropcode· Code & commands- Work product
WP09· Files & folders- Conditions
long· Long-horizonmultiapp· Multi-app- Proposed difficulty
- 4 / 5
Proposed verifier
120 files present; each file's ASR transcript (Whisper, fixed model) WER<=20% vs its prompt; the chosen take matches the gold last-fluent take (start within 500 ms); padding 100-250 ms; TextGrid text exact.
Proposed partial credit
15% 120 files named; 40% correct prompt per file; 25% last fluent take chosen; 20% padding and TextGrids
R08.10.01.04-T15 Measure jitter, shimmer and HNR on sustained vowels
Original instruction
input/sustained/ holds /a/ phonations from 12 speakers (3-5 s each). Per protocol input/voice_protocol.pdf select the most stable 1-s portion from the middle, then in Praat measure local jitter, local shimmer and HNR (cross-correlation, 75-500 Hz) and write output/voice.csv.
- Platforms
mac· macOSwin· Windows- Applications
- Praat
- Interfaces
desk· Desktop GUI- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
CL· CLI-sufficient- Interaction primitives
av· Audio/video over timeclick· Point & clickcode· Code & commands- Work product
WP02· Records/table- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Checker recomputes on the selected window (start/end columns required): window lies within the stable region per gold, and jitter, shimmer and HNR within 10% of gold measured on the gold window.
Proposed partial credit
30% windows within stable region; 70% 36 measures within tolerance (proportional)
R08.10.01.04-T16 Import a Praat TextGrid into ELAN and synchronize audio to video
Original instruction
input/interview.mp4 and a separate lavalier recording input/lav.wav start at different times; both contain a hand clap. Import input/lav.TextGrid into ELAN, link both media, set the audio offset so they are synchronized, shift annotations accordingly, and save output/interview.eaf.
- Platforms
mac· macOSwin· Windows- Applications
- ELAN · Praat
- Interfaces
desk· Desktop GUI- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
CL· CLI-sufficient- Interaction primitives
av· Audio/video over timedrag· Drag & dropmulti· Cross-app- Work product
WP10· Annotations- Conditions
multiapp· Multi-app- Proposed difficulty
- 3 / 5
Proposed verifier
EAF media descriptors include both files with TIME_ORIGIN offset within 20 ms of gold (2.36 s); imported tiers present and their times shifted consistently (max error 20 ms vs gold).
Proposed partial credit
30% TextGrid tiers imported; 40% offset within 20 ms; 30% annotation times aligned
R08.10.01.04-T17 Build a 7-step F0 continuum by PSOLA resynthesis in Praat
Original instruction
From input/yes_question.wav create 7 stimuli whose final-syllable F0 contour steps linearly from the original rise (question) to a fall ending at 110 Hz (statement), keeping duration and intensity unchanged. Use Manipulation/PitchTier editing and overlap-add resynthesis; save output/step1.wav..step7.wav.
- Platforms
mac· macOSwin· Windows- Applications
- Praat
- Interfaces
desk· Desktop GUI- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
CL· CLI-sufficient- Interaction primitives
drag· Drag & dropav· Audio/video over timecode· Code & commands- Work product
WP05· Time-based media- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
Checker measures F0 at 4 time points on the final syllable of each step: values linear across steps within 3 Hz of the gold endpoints-interpolated targets; duration within 5 ms of original; RMS within 1 dB.
Proposed partial credit
20% 7 files; 20% endpoints match targets; 40% linear spacing; 20% duration/intensity preserved
R08.10.01.04-T18 Compare two annotators in ELAN and build a consensus tier
Original instruction
input/child_speech.eaf has tiers 'utt_RA1' and 'utt_RA2' from two annotators. Use ELAN's Compare Annotators (modified kappa) and export output/agreement.txt; then create tier 'utt_consensus' that keeps overlapping segments using the earlier start and later end, and adds non-overlapping segments only if both are longer than 300 ms.
- Platforms
mac· macOSwin· Windows- Applications
- ELAN
- Interfaces
desk· Desktop GUI- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
CL· CLI-sufficient- Interaction primitives
click· Point & clickdrag· Drag & drop- Work product
WP10· Annotations- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
agreement.txt kappa value within 0.01 of gold run; utt_consensus annotations equal the rule-derived gold set (boundaries within 1 ms) and source tiers unchanged.
Proposed partial credit
30% agreement export; 55% consensus tier correct (proportional); 15% source tiers unchanged
R08.10.01.04-T19 Mark disfluencies in a 5-minute monologue on an Audacity label track
Original instruction
Open input/monologue.wav in Audacity. Add a label track with one label per disfluency, labelled 'UH', 'UM', 'REP' (repetition), 'REV' (revision) or 'PRO' (prolongation), each spanning the disfluent stretch. Export output/disfluency_labels.txt.
- Platforms
mac· macOSwin· Windows- Applications
- Audacity
- Interfaces
desk· Desktop GUI- GUI need · task setting
G2· GUI-essential- GUI need · ordinary access
CL· CLI-sufficient- Interaction primitives
av· Audio/video over timedrag· Drag & dropkey· Shortcuts- Work product
WP10· Annotations- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Event matching to gold (centre within 300 ms): detection F1>=0.75, type accuracy on matched events>=0.80. Audacity's mod-script-pipe scripting module is disabled and the VM has no shell or Python, so labels must be placed in Audacity; file writes by other processes are caught by file-access auditing.
Proposed partial credit
15% label export valid; 50% detection F1; 35% type accuracy
R08.10.01.04-T20 Find mislabeled vowel tokens by formant outliers and listening
Original instruction
input/vowel_tokens/ has 60 vowel WAVs with labels in input/labels.csv; 6 are mislabeled. Using Praat, plot the tokens in F1-F2 space, listen to suspects, and write output/mislabeled.csv with token id and the correct vowel label.
- Platforms
mac· macOSwin· Windows- Applications
- Praat
- Interfaces
desk· Desktop GUI- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
CL· CLI-sufficient- Interaction primitives
av· Audio/video over timevis· Visual judgmentcode· Code & commands- Work product
WP12· Answer/decision- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Set of flagged tokens compared with the 6 gold IDs (2 are acoustic near-neighbors only resolvable by listening); score = F1 of IDs x accuracy of corrected labels.
Proposed partial credit
60% ID F1; 40% correct replacement labels
Mapped research references
Existing benchmark-to-class links from the research catalog. They are not record-level annotations, executions or product endorsements.