Estudias medicina
Statement
−6.6
Question
+2.1
Three numbers
−4.75ststatement-final12 statements, all negative
−1.90stquestion-final12 questions
+2.15stsame sentence, question minus statement11/12 pairs agree
“Final” = the median pitch of the last 120 ms of voicing, relative to that sentence's own median pitch, in semitones. Measuring relative to the sentence's own median (rather than absolute Hz) is necessary because voices and sentences differ in overall pitch — what matters is how far the ending drops.
Twelve minimal pairs
One row per sentence. The dark dot is the statement, the green dot the question; the gap between them is what a single punctuation mark does.
Eleven of twelve rows point the same way: the question dot is to the right. The only pair that nearly overlaps is Quieres café — the method section says why.
Six pairs, heard in pairs
The same sentence with one punctuation mark changed. Play the statement first, then the question, and listen to how the last word lands.
Comes conmigo
Statement
−3.9
Question
+0.5
Piensas volver
Statement
−4.2
Question
−0.2
Trabajas en Madrid
Statement
−3.7
Question
−0.1
Llegas temprano
Statement
−4.0
Question
−0.7
Quieres café
Statement
−7.5
Question
−7.6
Three rules
Spanish intonation is clearer than English: the sentence type decides the final movement, with almost no exceptions.
01
Statements: the ending must fall into the low range
This is the signal for “I have finished”. All 12 statements here end 3.5–7.6 semitones below their own median. This is where Mandarin speakers slip: a Mandarin statement only dips slightly and stays essentially level, so a literal transfer sounds unfinished — or tentative.
02
Yes/no questions (¿…?): the ending does not fall, or rises
Adding a question mark lifts the ending 2 semitones relative to the statement version (11 of 12 pairs agree). Note that a Spanish yes/no question does not float upward throughout — the body still declines; only the last word lifts.
03
Wh-questions (¿Dónde…?): high start, low ending
Questions with a question word fall at the end in the standard description (like statements); the difference is at the start, where the question word carries the peak. This page gives no examples of that type — our synthetic voice reads ¿Dónde…? with a final rise too, which contradicts the standard. Using it as an example would teach the wrong thing.
In one line: statements fall; yes/no questions lift only the last word. A flat statement is a worse error than a wrong consonant — a wrong consonant sounds like an accent, a wrong contour changes whether you are stating or asking.
Where these numbers come from
Why synthetic: this page needs two versions of the same sentence. The real corpus cannot supply that — all 145 human recordings are declarative sentences from novels, with zero questions and zero exclamations. Changing only the punctuation keeps the words, the voice and the tempo fixed, which is a cleaner design than any human corpus could give.
Corpus: 12 minimal pairs (24 clips), one Spanish female voice, 5–9 syllables each. The question is generated from the statement by a text function (¿ + drop the full stop + ?), not written by hand a second time.
Measure: the median F0 of the last 120 ms of voicing, relative to the median F0 of all voiced frames in that sentence, in semitones (12·log₂). Not a linear regression over the final segment — that is what we tried first, with a max(350 ms, 25%) window, and only 4 of 6 pairs agreed: the 350 ms window averaged away the question rise, which in the pitch track happens only in the last ~150 ms. With the 120 ms window, 11 of 12 agree.
Result: statements −4.75 st (range −7.55 ~ −3.45), questions −1.90 st (range −7.57 ~ +2.13), paired difference median +2.15 st, 11/12 in the same direction.
Counter-example reported: 1 of 12 pairs shows almost no difference from the engine (Quieres café, Δ −0.02 st). That is why the page says “11/12” and not “all”. It is also why practice cannot consist of following this page's audio — you have to produce the fall yourself.
Engine limitation (stated on purpose): in our clips the wh-question rises at the end just like a yes/no question, while the standard description of Spanish ¿Dónde…? is a high start and a falling ending. That is the synthesiser treating every ¿? the same way, not a fact about Spanish. So this page gives no wh-question examples and no audio to imitate.
Measurements, not verdicts. Intonation is heard rather than noticed — native listeners usually cannot say what is wrong, they just feel they are being asked something.