Kategori: Voice & Speech Tools

Everyone with a podcast has been told to cut it into short vertical clips. The advice is
sound and the work is miserable: scrubbing a sixty-minute timeline looking for thirty good
seconds is an afternoon nobody has.

Automation cannot do the whole job, but it can do the tedious half.

What a machine can actually detect

Without understanding the words, the one signal a machine can read reliably is
energy — how loud each moment is relative to the rest.

That turns out to correlate with interesting moments more often than you would expect.
People get louder when they laugh, when they interrupt, when they emphasise, when they get
annoyed, and when they finally say the thing they have been circling for ten minutes. Energy
finds all of those.

It also finds coughs, microphone bumps and someone knocking a mug over. Which is the point
of the caveat further down.

How the moments get picked

The recording is measured one second at a time, so every moment has its own level rather
than one average for the whole file. That distinction matters more than it sounds: a summary
figure for a sixty-minute episode tells you nothing about where anything happened.

From there:

• The loudest moments are taken in order.
• A minimum gap is enforced so two clips never cover the same passage. Without it, one
  animated thirty-second exchange produces five nearly identical clips.
• Each clip is built around its peak with lead-in and follow-through — a few seconds
  before, the rest after — so the listener hears what led up to the moment rather than
  landing in the middle of it.
• Short fades at both ends stop a clip starting with a click or cutting off mid-word.

Why fifteen seconds

Long enough to carry one complete thought, short enough for every vertical video platform,
and short enough that a viewer will finish it. Every clip being the same length also makes them
easy to batch: same template, same captions, same publishing routine.

The honest limitation

Energy tells you where something happened. It cannot tell you whether it was
interesting. A cough, a laugh at a terrible joke and a genuinely brilliant one-liner look
identical to a level meter.

So treat the output as a shortlist, not a publishing queue. Ten candidate clips take seconds
to produce and about three minutes to listen through. Two or three will be usable. That is a
good trade against an afternoon of scrubbing — but it is a shortlist, and the listening
step is not optional.

A related honesty point about sensitivity controls: on a lively episode with plenty of loud
moments, a strict setting and a loose one pick the same top ten, because the loudest moments are
the loudest either way. The control earns its place on quiet or uneven recordings, where a
strict setting comes back with two clips and a looser one finds eight.

What to do with the clips

You get audio, not video — the visual layer is still yours to add. The usual routes:

• A waveform or static image with burned-in captions, assembled in whatever editor you
  already use.
• A talking-head crop, if you filmed the session.
• The timestamps alone, used as a map back into your video project.

That last one is underrated. Even if you never publish a single audio clip, knowing that the
five most animated moments of the episode were at 12:04, 19:30, 27:15, 41:52 and 48:07 saves
the entire search.

Where it fits in the routine

Publish the episode first, then mine it. Clean up any steady background noise, master the
audio to a consistent loudness, publish, then run the finished file through clip cutting and
work from the shortlist. The clips inherit whatever quality the master has, so doing it in that
order means you are not cutting up a noisy file.

Back catalogues are where this pays off most. Fifty old episodes nobody has time to mine
become fifty shortlists, and the good moments in them are just as good as they were the week
they went out.

Speech engines read ordinary sentences well. They fail on the parts that are not
sentences: figures, abbreviations, links, addresses. The failures are quiet — nothing
errors, you just get a file with a wrong word in the middle of it — and you only find
them by listening to the whole thing.

Ten minutes of preparation removes almost all of them.

Numbers are the main event

A bare figure forces the engine to guess how to say it, and the guess depends on
surrounding context in ways you cannot predict. In our own testing a four-digit number
followed by a hyphenated word came out as something that was not a number at all.

Writing it out in words removes the guess. 5070 becomes “five thousand
seventy” and there is nothing left to interpret. This is the single highest-value edit you
can make to a script.

Thousands and decimals are language-specific, and they are opposites

This trips people up constantly, because the two conventions are mirror images:

• English: 1,180 is one thousand one hundred eighty;
  3.14 is a decimal.
• Turkish and much of Europe: 1.180 is one thousand one
  hundred eighty; 3,14 is a decimal.

Read a European price with English rules and a figure meaning three and a bit becomes three
hundred and fourteen. The two languages also say decimals differently: English reads
the digits one at a time (“three point one four”), while Turkish reads the fraction as a whole
number.

Leave codes alone

Not every digit should be converted. MP3, H264 and model numbers
are read correctly as they are, and spelling them out makes things worse. The rule that works:
convert digits that stand alone, leave digits attached to letters.

The rest of the checklist

• URLs and e-mail addresses. Read aloud, they are gibberish, and every
  character is billed. Replace them with a short phrase or remove them.
• Repeated whitespace and stray line breaks. Harmless to meaning, but they
  count toward your quota and can affect phrasing.
• Curly quotes and long dashes. Normalise them. Some engines pause oddly
  on typographic punctuation.
• Bracketed notes and stage directions. If it is not meant to be spoken,
  take it out — it will be spoken.

Split before you send, not after it fails

Most engines cap a single request at around 5,000 characters. A long chapter has to be
split, and where you split matters: break on sentence boundaries, never mid-sentence, or the
join will be audible.

Knowing the request count in advance is also how you plan the cost. Speech is billed per
character, so the character count of your cleaned text is the price. Clean first,
count second, and you know what an audiobook chapter costs before you commit to it.

How long will it actually be?

Words divided by a speaking rate. 180 words per minute is a typical narration pace;
audiobooks usually run slower, conversational podcasts faster. A 9,000-word chapter is
roughly 50 minutes at a normal pace — useful to know before you find out the hard way
that it is too long for one episode.

Readability, and one caveat

Readability scores are worth a glance, with a warning: the well-known formulas are
calibrated per language. Flesch Reading Ease is built on English syllable patterns and returns
a meaningless figure for an agglutinative language like Turkish, which has its own adaptation
(the Ateşman index). A tool that applies the English formula to any text is giving you a number
that looks authoritative and means nothing.

Why do this in the browser

Scripts are often unpublished manuscripts, client work or internal documents. Text
preparation is pure string manipulation — there is no reason for any of it to leave your
machine. If a tool uploads your draft to a server to count its characters, that is a choice it
made for its own convenience, not a technical requirement.

The order that works

1. Paste the text and look at the character count.
2. Convert numbers to words in the right language.
3. Clean URLs, addresses and stray whitespace.
4. Check the request count and the longest segment.
5. Send it to synthesis.
6. Master the result before publishing.

Steps two and three take a minute and remove the class of error that is hardest to find
later.

There is a useful line running through every noise problem: is the noise
steady or not?

Steady noise — air conditioning, a fan, a computer, tape hiss, mains hum — has a
consistent fingerprint. Measure it during a quiet moment and you can subtract it from the whole
recording. This works well, and it is what noise removal actually is.

Everything else — a door, a dog, traffic, someone talking over you — is not
steady. There is no fingerprint to subtract. No filter of this kind removes it, and any tool
promising otherwise is selling you something.

Why fixed settings do not work

Every room is noisy at a different level. A denoiser set to expect noise at one level will
do almost nothing in a quieter room and destroy the audio in a louder one.

We measured this directly. With a fixed setting, the noise floor of our test file dropped
by one to two decibels — effectively nothing. Measuring the file’s actual noise floor
first and aiming the filter at that produced between 8 and 44 dB of reduction
depending on the strength chosen. Same filter, same file; the only difference was whether the
tool looked before it acted.

So the first question to ask any noise remover is: does it measure my file, or does it apply
the same setting to everything?

How much is too much

Here are real measurements from four strength settings on the same recording:

• Soft: 8.6 dB of noise removed
• Normal: 11.8 dB
• Aggressive: 22.1 dB
• Brutal: 43.5 dB

Across all four, the speech level moved by less than one decibel. That is the number to
watch: if a denoiser is making your voice quieter, it is removing the wrong
thing.

More is not automatically better. Very strong subtraction starts to thin out breaths and the
quiet ends of words, which is how audio gets that hollow, underwater quality. Start in the
middle, listen to a short preview, and only go further if you can hear that you need to.

The mistake worth avoiding: normalising afterwards

It seems obvious — clean the file, then bring it up to publishing loudness in the same
pass. We measured what that costs and it is worse than it looks.

Signal-to-noise improved by 11.1 dB with normalisation switched off, and only 2.2 dB with it
on. Normalising applies gain to everything left in the file, including whatever noise
survived, and its dynamic stage lifts the quiet passages hardest — which is exactly the
part you just cleaned.

The right order is two separate steps: remove the noise, then normalise the cleaned file
with a proper two-pass loudness tool. Same tools, same effort, dramatically better result.

Do not forget the low end

A high-pass filter around 80 Hz costs nothing and removes desk thumps, chair creaks,
handling noise and mains hum. Almost no speech content lives below 80 Hz, and that rumble
takes up headroom you would rather spend on the voice. Turn it on unless you have a specific
reason not to.

What about gating?

A gate silences anything below a threshold. Between sentences that sounds wonderfully clean;
at the end of words it chops the natural tail off consonants and breaths, and the result is
choppy in a way that is hard to unhear. Leave it off unless you have listened and decided you
want it.

Fix it at the source when you can

Every decibel you do not record is a decibel you never have to remove. Turn the air
conditioning off during takes. Move the microphone away from the laptop. Get closer to the
microphone so your voice sits further above the room. Five minutes of setup beats any amount
of processing.

When that is not possible — a remote guest with a noisy laptop, an archive recording,
a room you do not control — measure, remove, then normalise. In that order.

Five formats cover almost everything you will ever be handed. Knowing what each is for
takes about five minutes and saves a lot of pointless conversion.

The only distinction that matters

Lossless formats keep every sample of the original. WAV stores them
uncompressed; FLAC compresses them without discarding anything, ending up roughly half the
size for identical audio.

Lossy formats throw away detail the ear is least likely to miss, which is
why they are five to ten times smaller. MP3, M4A and OGG are all lossy.

Everything else follows from that one line.

The five, briefly

MP3

Plays on absolutely everything, and that universality is the whole reason to choose it.
192 kbps is the sensible default for speech and music alike. 128 kbps is fine for plain
narration and noticeably smaller. 320 kbps only helps when the source is already high
quality — it cannot restore what a bad recording never had.

WAV

Uncompressed, lossless, and enormous: roughly 10 MB per minute in stereo. Exactly right
when a file is going back into an editor for more work. Wrong for sharing, and wrong for
anything you are emailing.

M4A

AAC audio in an MP4 container. Slightly better than MP3 at the same bitrate and the
default across Apple devices, which is where most voice memos come from. Support is very
good but not quite MP3-universal.

OGG

Vorbis, a genuinely open format with true variable bitrate — it spends bits where the
audio needs them rather than at a fixed rate. Excellent quality per megabyte. The catch is
compatibility: plenty of hardware players and upload forms still refuse it.

FLAC

Lossless but compressed. About half the size of WAV with bit-identical audio. This is the
right archive format: keep your masters as FLAC and export lossy copies for publishing.

The conversion that is always a waste

Converting a lossy file to a lossless one does not bring back what was discarded. An MP3
turned into FLAC is a larger file containing exactly the same audio, with the same artefacts,
now stored inefficiently.

Convert to lossless only when your source is lossless. If someone sends you an MP3
and asks for “the WAV”, what they will get is an MP3 in a bigger box.

Generation loss is real but often overstated

Every lossy-to-lossy conversion decodes and re-encodes, losing a little more each time. One
conversion at a sensible bitrate is inaudible on speech. Five conversions at 128 kbps are
not.

The rule that avoids the problem: keep the original recording, and convert from it each
time rather than from an export of an export.

Mono or stereo?

For a single narrator or an interview recorded on one microphone, mono halves the file
with no audible loss — there is no stereo information to preserve. Keep stereo when the
recording genuinely has left and right content: music, ambience, or hosts recorded on separate
channels and panned.

Do your tags survive?

They should. Title, artist, album and comments carry across all five formats when the
converter asks for it. One detail catches people out: MP3 tags should be written as ID3v2.3
rather than the newer 2.4, because a surprising number of older players and car stereos
cannot read 2.4 and will show your episode as “Unknown”.

Quick answers

• Publishing a podcast? MP3 at 192 kbps.
• Sending audio to an editor? WAV, or FLAC if size matters.
• Archiving masters? FLAC.
• An upload form refused your file? MP3 — it is the format nothing
  rejects.
• Voice memo that will not open on Windows? M4A to MP3.

And if the file is simply too big rather than the wrong type, the answer is a lower bitrate
or mono, not a different container.

A listener switches from someone else’s show to yours and reaches for the volume knob.
That is not a content problem, it is a loudness problem, and it has a precise fix that takes
about a minute.

Loudness is not volume

Peak level tells you how close the loudest instant comes to distortion. It says nothing
about how loud something sounds. A recording can peak near the ceiling and still feel
quiet, if the loud moments are rare and everything between them sits low.

LUFS measures perceived loudness across the whole programme, weighted roughly the way
human hearing works. That is why platforms specify LUFS and not peak: it is the number that
predicts whether a listener will reach for the knob.

The numbers worth knowing

• −16 LUFS — the usual target for stereo podcast delivery.
• −14 LUFS — where YouTube and the music streaming services
  normalise to.
• True peak — a separate ceiling, typically −1.5 dBTP for
  podcast and −1.0 dBTP elsewhere, that the audio must not exceed after conversion.

Publish louder than the target and the platform turns you down anyway. Publish quieter and
some platforms leave you quiet. Hitting the number is the whole job.

Why true peak sometimes wins

This is the part most guides skip, and it is the reason a masterer sometimes tells you
“that is as loud as it goes”.

Loudness only rises by adding gain, and gain raises the peaks by exactly the same amount.
If your recording is quiet and spiky — a big dynamic range between the loudest
consonants and the average level — reaching −14 LUFS can require more gain than
the peak ceiling allows. Something has to give: either the ceiling breaks and the audio
distorts, or you stop short of the target.

Stopping short is the correct answer. A file at −15 LUFS that respects the ceiling
sounds fine. A file at −14 with clipped peaks sounds broken, and no platform will fix
it for you.

Two-pass normalisation, and why one pass is not enough

A single-pass loudness filter guesses as it goes. On uneven material it drifts, and the
finished file lands somewhere near the target rather than on it.

Two-pass measures the whole file first, then applies a correction calculated from those
measurements. The difference is not academic: on quiet or uneven recordings a single pass can
miss by several decibels, which is exactly the error you were trying to remove.

Better still is to measure the result afterwards and correct once more if it
landed outside tolerance. That is the difference between “we applied a loudness filter” and
“the file measures −16.1 LUFS”, which is a claim you can check.

The rest of the mastering chain

Loudness is one step of four, and the order matters.

1. High-pass first. Cutting below about 80 Hz removes desk thumps, chair
   creaks and mains hum. Do it before measuring, or that rumble eats headroom you needed.
2. Trim the silence. Recorders run before you speak and after you stop.
   Cut the dead air at the ends, and leave the pauses inside the conversation alone —
   they are the rhythm of the thing.
3. Normalise. Measure, then apply.
4. Fades last. A fade applied before normalisation gets partly undone by
   it. Half a second in and a second out is a sensible default.

What mastering cannot fix

It sets the level, tidies the ends and protects the peaks. It does not remove room echo,
repair clipping already baked into the recording, balance two hosts recorded on separate
tracks, or make a bad microphone sound good. Mastering is the last step, not a rescue —
and if the recording has steady background noise, clean that before you normalise,
not after.

A workflow that holds up

Record with headroom, aiming for peaks around −6 dB rather than as loud as possible.
Edit. Remove steady noise if there is any. Then master: high-pass, trim, normalise to
−16 LUFS for the podcast feed, fade. If the same audio also goes to YouTube, run it
again at −14 rather than uploading the podcast master and hoping.

Check the output number. If a tool cannot tell you what the finished file measures, it is
asking you to take its word for it.