An investor listening to a pitch decides quickly whether it is worth continuing, and the first thing they want to know is what you have built, not how well you speak. Your voice matters, but it matters after the content: people who have seen many live pitches, like the Y Combinator partners in the essay you will read shortly, put "what you do" at the top of the list and everything else much lower down. Pitch Coach Audio is a tool on this site that gives you automatic feedback on a pitch rehearsal: three measures computed on the text (pace, filler words, estimated pauses) and a qualitative reading on five dimensions. This guide explains what those measures mean, where they go wrong and how to use them without letting a number convince you of something.
One clarification right away: the tool analyses text. Audio is only there to produce the transcript, and the tone of your voice cannot be derived from text. When you read "Energy" in the results, that score comes from things that can be measured on the text and on the duration, not from an analysis of intonation. Keep this in mind throughout the article, because it changes the right way to read the results.
What people who have seen many live pitches say
Before measuring anything it helps to know what is worth measuring. The most useful text I have read on the subject is "How to Present to Investors" by Paul Graham, written in 2006 (revised in 2007 and 2010) after a rehearsal day for Y Combinator's Demo Days, when the four partners realised they were repeating the same things to founders at every batch. The essay starts from this: with ten minutes it is best to aim at two goals only, explaining what you do and why users will want it. Everything else competes with those two for a few minutes.
From there the essay draws a series of very concrete tips, which I report in my own words. Say what you do in the very first sentence, with no preamble. A narrow, slightly reductive description is better than a broad, vague one, because the vague one turns into noise (the essay's own example of boilerplate, "a system to allow people to collaboratively leverage the value of information", tells nobody anything). Competitors and CVs go on one quick slide, at the end. The business model gets a mention, but without going into detail: the author observes that whatever you have at this stage is probably wrong. If you have data, give a few numbers specific to the company, four or five at most, and leave the market size alone. Slides, if you can manage it, stay under twenty words and are not read aloud.
On the way of speaking, which is the tool's territory, the essay says two things that are worth keeping together. The first is to speak slowly and clearly, facing the audience, in a voice different from ordinary conversation: the author reports a rule used by actors, mentioned by one founder, according to which if you feel you are speaking too slowly you are speaking at about the right speed. The second is not to overdo the delivery, because an announcer's tone tires the audience: you have to sound natural. I also report a caution from the author himself: the section on the demo is declared obsolete in the text for YC Demo Days, because presentations are now too short to include one.
An invented example shows the difference between a weak opening and a strong one. Suppose a startup that sells practice-management software for dental clinics. Weak opening: "In recent years the digitalisation of the healthcare sector has been changing the habits of professionals." Strong opening: "We are three founders and we built software that sends reminders to dental clinics, so they stop losing appointments." The second sentence says who you are, what you did and for whom, in one line. Neither sentence contains market data, and indeed the second does not need any.
What the tool does, and what it does not
The tool offers three ways to provide the pitch: record from the microphone (up to 60 seconds), upload an audio file or paste the text. In practice the first and the last are the ones that count. In a recording the transcript is produced by the browser's speech recognition; the MDN documentation on the SpeechRecognition interface warns that in some browsers, Chrome included, recognition uses a web service, so the audio goes to that service. The site does not receive the audio: it receives the transcribed text. But the audio may leave your device towards the browser vendor, and if the pitch contains confidential information that is worth knowing before you record.
Uploading an audio file, right now, transcribes nothing: the page accepts mp3, wav and webm up to 5 MB, but then asks you to paste the text by hand in the other mode. If you already have a transcript, pasting the text is therefore the most direct route, with a limit of 8,000 characters for the analysis.
There are two outputs. The first is the metrics, computed immediately in the browser:
- words per minute, with a label ("Troppo lento", too slow, below 100; "Ottimale", optimal, up to 140; "Veloce", fast, up to 180; "Troppo veloce", too fast, beyond that);
- the number of filler words, counted against a fixed list of Italian expressions (tipo, ehm, ah, insomma, nel senso, come dire, appunto, praticamente, diciamo, quindi);
- estimated pauses, meaning gaps longer than 800 milliseconds between one recognised segment and the next;
- the duration of the pitch.
These four items are not on the same level. Duration and words are facts, pace is a ratio between the two, while fillers and pauses are estimates that depend on how the browser splits the speech and on a word list chosen in advance. Read them as clues to compare from one rehearsal to the next, not as grades.
The second output is the qualitative reading: an overall score and five dimensions (Clarity, Value Proposition, Problem-Solution Fit, Ask, Energy), each with a mark from 1 to 10 and a one-sentence rationale, plus strengths and areas for improvement. It is produced by an AI engine that runs on the site's server: since 12 November 2026 it is a local engine, with a job queue, and that is why the wait is longer than for an immediate answer. If the engine does not respond, the tool does not leave you without a result: it computes the scores from the metrics alone and tells you so. In that case Value Proposition, Problem-Solution Fit and Ask are worth 5 with the rationale "non valutabile senza AI" (not assessable without AI): they are not a judgement on your pitch, they are a placeholder. If you see three fives in a row with that sentence, the content has not been read.
The metrics, recomputed from scratch
To understand how far to trust the numbers it helps to rebuild them. The program below is not the tool's code: it is a replica written from scratch with the same definitions (words divided by duration times sixty, a list of fillers, a pause over 800 ms), applied to an invented 42-second pitch, so that you can look inside the three measures. I ran it twice with Python 3 and got the same output. The words per minute are 51 divided by 42 times 60, that is 72.86, rounded to 73: if you redo the sum by hand it matches. The pitch text stays in Italian on purpose, because the tool's filler list is Italian: "praticamente" is basically, "quindi" is so or therefore, "tipo" is like, "ehm" is um.
import re
# INVENTED pitch for the test (no real company), known duration: 42 seconds.
# The text stays in Italian on purpose: the tool's filler list is Italian.
TRANSCRIPT = (
"Ciao, siamo tre fondatori e abbiamo costruito un gestionale per studi dentistici. "
"Praticamente oggi uno studio perde un appuntamento su dieci perché nessuno manda il promemoria. "
"Quindi il nostro software lo manda da solo, tipo via messaggio, e riempie la poltrona libera. "
"Ehm, chiediamo duecentomila euro per sei mesi di lavoro."
)
DURATION_S = 42
# Arrival times of the recognised segments, in milliseconds from the start.
SEGMENT_TIMES_MS = [0, 3200, 7400, 9000, 12500]
FILLERS = ["tipo", "ehm", "ah", "insomma", "nel senso", "come dire",
"appunto", "praticamente", "diciamo", "quindi"]
PAUSE_MS = 800
def word_count(text):
return len(text.split())
def wpm(text, duration_s):
return int(word_count(text) / duration_s * 60 + 0.5) if duration_s > 0 else 0
def count_fillers(text):
pattern = r"\b(" + "|".join(FILLERS) + r")\b"
found = [t.lower() for t in re.findall(pattern, text, flags=re.IGNORECASE)]
return {w: found.count(w) for w in sorted(set(found))}
def pauses(times_ms):
return sum(1 for a, b in zip(times_ms, times_ms[1:]) if b - a > PAUSE_MS)
print("words :", word_count(TRANSCRIPT))
print("words per minute :", wpm(TRANSCRIPT, DURATION_S))
print("fillers :", count_fillers(TRANSCRIPT))
print("pauses over 800 ms:", pauses(SEGMENT_TIMES_MS))
# The same words, used in their normal sense: the count cannot tell.
sentence = "Il costo sale, quindi il prezzo sale. Appunto: il margine, quindi, scende."
print("false positives :", count_fillers(sentence))
# Output: words : 51
# words per minute : 73
# fillers : {'ehm': 1, 'praticamente': 1, 'quindi': 1, 'tipo': 1}
# pauses over 800 ms: 4
# false positives : {'appunto': 1, 'quindi': 2}
The result says two things. The first: 73 words per minute is a slow pace, in the "Troppo lento" band of the tool's label, and indeed the invented pitch has long pauses between segments (four gaps above 800 ms). It is not necessarily a flaw: it recalls the actors' advice reported in the essay, and speaking slowly is, according to that text, the right direction. The label thresholds are a setting of the tool, not a law: no source I have read fixes the "optimal" pace of a pitch in words per minute, and that is why I am not giving you a number to chase.
The second thing is in the last line. The filler list counts "quindi" and "appunto" (so, precisely) even when they are ordinary connectives: "il costo sale, quindi il prezzo sale" (the cost rises, so the price rises) contains two legitimate "quindi" and marks them as tics. On a long pitch the error weighs little, but on a 40-second rehearsal it can shift the count by a couple of units. Before deciding that you speak "in fillers", reread the transcript and look at which words were counted.
That leaves the "paste text" mode. Here the page does not know the duration and estimates it at 130 words per minute, as the hint on the tool's page also states. The next program repeats the calculation for every length from 10 to 1,300 words and shows what that choice implies for the words per minute reported.
# In "paste text" mode the page does not know the duration and estimates it at 130 words per minute.
# What happens, then, to the words per minute it shows?
def round_half_up(x): # half rounds up, like Math.round
return int(x + 0.5)
def wpm(n_words, duration_s):
return round_half_up(n_words / duration_s * 60)
results = {}
for n in range(10, 1301):
estimated_duration = round_half_up(n / 130 * 60) # the page's estimate
results[n] = wpm(n, estimated_duration)
values = sorted(set(results.values()))
print("lengths tried :", len(results), "(from 10 to 1300 words)")
print("min and max :", min(values), max(values))
exact = sum(1 for v in results.values() if v == 130)
print("exactly 130 :", exact, "out of", len(results))
# Output: lengths tried : 1291 (from 10 to 1300 words)
# min and max : 120 140
# exactly 130 : 1159 out of 1291
The value is always around 130: out of 1,291 lengths tried, 1,159 give exactly 130 and the others oscillate between 120 and 140 because of the rounding of the duration (confirmed by the formula: words divided by estimated duration, where the duration is words divided by 130 times sixty, returns 130 apart from the rounding). In other words, when you paste a text the pace shown measures nothing about your speaking: it is the page's own assumption reflected back on the screen. The number only makes sense if you really record with the microphone, or if you enter the duration yourself in another tool.
How to use the feedback without being led by it
Automatic feedback is useful when you use it to compare yourself with the previous rehearsal, and less so when you take it as a verdict. A sensible routine proceeds in this order, because each step presupposes the previous one: first the message, then the form, then the comparison with real people.
- Write the first sentence of the pitch and check that it says what you do, for whom and what problem it solves, with no preamble.
- Record or paste a rehearsal and read the Value Proposition dimension and the areas for improvement first, not the overall score.
- Check the metrics: pace, fillers really counted, pauses. Change one thing at a time, so that the next rehearsal tells you something.
- Repeat the pitch in front of a person who does not know your sector and ask them to tell you in their own words what you do.
The order matters because optimising the pace of a vague message means performing better something that cannot be understood. The fourth step is the one the tool cannot do: Graham's essay observes that the people listening are often not very technical, and only such a person will tell you where you lose them. The title of this article speaks of ninety days: no source I have read says how long it takes for a pitch to improve, so I am not promising you a duration, only a method.
Frequently asked questions
How long should a pitch be?
It depends on the format. In Paul Graham's essay the Y Combinator Demo Day gives each startup ten minutes and the author advises concentrating them on two goals, explaining what you do and why users will want it. That text was last revised in 2010, and formats have changed since: there is no right length in the absolute, there is the one the event gives you.
How many words per minute should I speak?
There is no threshold with a source. The tool's label calls "Ottimale" (optimal) the stretch up to 140 words per minute, but that is a setting, and the essay cited advises erring on the side of slowness. Use the measure to compare two rehearsals of your own, not to reach a number.
Is the audio I record sent to the site?
The site receives only the text of the transcript, not the audio file. But the transcript is produced by the browser, and according to the MDN documentation in some browsers, Chrome among them, speech recognition sends the audio to a web service. If the pitch contains confidential data, paste the text instead of recording.
Can I upload an audio file I have already recorded?
You can select mp3, wav or webm up to 5 MB, but the automatic transcription of the file is not performed: the tool asks you to paste the text. For now the file only serves as a reminder of the rehearsal.
What you can do now
You can separate what the tool measures (words, duration, fillers, estimated pauses) from what the model estimates (the five dimensions), and you can recognise the cases in which a number says nothing: the pace in "paste text" mode, the three fives in a row of the fallback without AI, the "quindi" counted as tics. You have a four-step routine for improving a pitch starting from the message. For the next step, when the pitch has to become a negotiation, read the guide to the term sheet; for the price and VAT of the product you present, the one on pricing; for the first hires, the one on hiring in a European startup. The tool is at the Pitch Coach Audio page.
Sources
- Paul Graham, "How to Present to Investors" (August 2006, rev. April 2007 and September 2010) - read in full on 2026-11-13: priority of the message, opening, slides, business model, numbers, pace of speaking.
- MDN Web Docs, SpeechRecognition interface - read on 2026-11-13: warning about recognition based on a web service in some browsers; limited availability of the API.
- Behaviour of the tool: on 2026-11-13 I read the description in the tools registry and the page code (metrics, label thresholds, 130 words per minute in "paste text" mode, fallback without AI).

Commenti
Caricamento commenti...
Sign in to comment