Episode 6485 custom

Choosing Audio Bitrates and Settings for Voice

Fullband voice encoding tops out at 16 kilobits per second mono in Opus, making high bitrates mostly redundant for speech.

Nov 7 21:49 Transcript

With

0:00

Episode Details

Episode ID: MWP-6687
Published: Nov 7
Duration: 21:49
Audio: Direct link
Pipeline: V5.26
TTS Engine: chatterbox-regular
Script Writing Agent: GLM-5.3
Topics: speech-recognition music audio-equipment

AI-Generated Content: This podcast is created using AI personas. Please verify any important information independently.

The prompt

The prompt that started this episode, as sent in (production notes for the hosts removed).

We've shared plenty of general thoughts on audio for voice notes and dictation, but this time I want to ground it in an actual app on an actual phone. Good voice recording apps on a phone expose a whole set of options. Everyone probably opens them up and thinks there has to be some good combination, but who knows what's best? I want a firm set of recommendations for real scenarios, based on my hardware: a OnePlus Android phone running the ASR Voice Recorder app.
I think most people record at an unnecessarily high fidelity, or don't know how to choose between the recording formats. Long recordings of more than an hour make that matter, because larger files risk hitting API limits.
Here are the options I have. Recording format: WAV, MP3, Opus, FLAC, OGG, M4A and AMR. I tend to go for Opus for voice, because I've always been under the impression that it's the best balance of size and performance, and I only use MP3 for general compatibility when a particular API needs it. Is that correct, and what are the other options for? I don't even know what AMR is.
Bitrate runs from 6 all the way up to 320 kilobits per second. I'm currently at 32, which is what Claude told me to go to, down from my previous setting of 64. Sample rate options are 8, 12, 16, 24 and 48 kilohertz. Audio channels default to stereo, and I commonly put that down to mono. There are also noise cancellation and automatic gain settings, which are slightly away from the core technical focus. And there's microphone zoom, which goes from 0 to 1 and from 0 to minus 1. I don't know exactly what it's doing. Is it a real hardware feature exposed on a phone with an array of microphones? I'm going to try 0.2 to get a little more directional.
That's my voice note profile described in full, and my podcast profile is those same settings changed up a little. So how would you set these for three scenarios on this same phone: recording an important meeting with your lawyer, a voice note to yourself that has to be good to listen to and also transcribe well, and recording a podcast? I'd love the specific numbers you'd pick for each.

Navigating recording app settings often leads to unnecessarily inflated audio files. Opus stands out as the ideal voice codec, combining speech and transform encoders to deliver fullband speech at just sixteen kilobits per second mono. While uncompressed WAV or lossless FLAC remains necessary to prevent generation loss in archival or forensic contexts, voice notes and podcasts achieve optimal efficiency around twenty-four kilobits per second mono. Legacy formats like MP3 persist purely for ubiquitous API compatibility, while 1999-era AMR offers little utility today. Additionally, the mysterious microphone zoom slider maps directly to the Android platform's setPreferredMicrophoneFieldDimension method. Originally created to align audio beamforming with camera video zoom, this setting merely requests hardware-level directional focus that device manufacturers can silently ignore.

Sources

What the research for this episode read before the script was written. Primary sources first.

Mentions

3GPP Mobile telecommunications standards body
AAC Advanced Audio Codec, common in web
AMR Legacy 3GPP telephone speech codec
ASR Voice Recorder Android recording app with format menu
FLAC Lossless audio codec for high-fidelity
OnePlus Android phone used for recording
Opus Modern audio codec, efficient at low bitrates
setPreferredMicrophoneFieldDimension Android AudioRecord API for mic beamforming
Vorbis Open codec Opus superseded
Xiph.Org Maintains Opus, FLAC, Vorbis guidance

Downloads

Episode Audio

Download the full episode as an MP3 file

Download MP3

Transcript (TXT)

Plain text transcript file

Episode Book (PDF)

The episode's record — date, duration, models, sources — with the full transcript

#6485: Choosing Audio Bitrates and Settings for Voice

What if I told you most of a recording app's settings menu is decoration, and the one slider nobody can explain is a camera feature that leaked into audio? That's roughly where Daniel has landed. He's got a OnePlus running ASR Voice Recorder, a menu full of formats and bitrates and sample rates, a microphone zoom slider he doesn't understand, and he wants firm numbers for three situations: an important meeting with a lawyer, a voice note that has to be good to listen to and transcribe well, and a podcast.

And he's already halfway there on his own. He dropped Opus from sixty-four kilobits a second to thirty-two, on the friendly AI down the road's advice, and he's asking whether that's right.

He suspects, and I quote, that "most people record at an unnecessarily high fidelity." He's correct, and the arithmetic backs him, but before the numbers, the formats. Because the format choice constrains everything downstream.

Seven formats in his list. WAV, MP3, Opus, FLAC, OGG, M4A, AMR. Small correction to the menu itself first: the app's own site lists an eighth, thirty-two bit WAV, which Daniel's list doesn't have. Worth checking his build, though for voice it changes nothing.

An eighth format nobody asked for. The menu grows, it never shrinks.

Never. Give each one a line. WAV is uncompressed PCM, the raw numbers off the microphone, biggest file on the menu, your archival and forensic master. Wasteful for a voice note, correct when the recording might end up in front of someone official. FLAC is lossless compression, so it's bit-identical to the WAV but typically about half the size, and it's the one I'd actually pick when lossless matters. Xiph's guidance on archiving is blunt: use a lossless format to prevent generation loss.

And generation loss is the thing people never see coming, right? Walk me through it.

It's the photocopy problem. You photocopy a page, it's fine. You photocopy the photocopy, then photocopy that, and by the fifth generation the text is fuzzy. Every lossy encode is a photocopy. You record to MP3, your editor re-encodes it when you trim the dead air off the front, your transcription service maybe re-encodes again, and each pass throws away a little more. Lossless breaks that chain, because the copy is exact, forever. FLAC is the page in the vault.

MP3, I assume, survives on pure ubiquity.

It's a odd format under the hood, raw audio coding without a container, with the tags bolted on afterwards as a hack. And that hackishness is why it plays everywhere. Every transcription API, every legacy player, every email client will take an MP3. Daniel's rule, MP3 only when a particular API needs it, is sound. Fallback, not default.

M4A?

Not a codec, a container. AAC audio inside an MPEG-4 wrapper. It's the Apple-friendly option, the one QuickTime and the Apple ecosystem handle natively. Fine, unremarkable. OGG is the open container, usually carrying Vorbis inside, and Vorbis is the codec Opus replaced. So OGG is the legacy open choice, superseded by its own successor. Then Opus, which we'll get to, and then...

AMR. The one he says he doesn't even know.

My favorite item on the list. AMR is a telephone codec. Standardized by 3GPP in October nineteen ninety-nine, encodes a band from two hundred to thirty-four hundred hertz, bitrates from four point seven five to twelve point two kilobits a second. Toll quality from seven point four. It's the format your old Nokia saved voice memos in.

So it's a museum piece sitting in a twenty-six-year-later app.

A museum piece with a docent. It exists for MMS and legacy voicemail systems and nothing else. Wikipedia's own caveat is that it's a speech format and unlikely to give ideal results for other audio. Never pick it. If you pick it, your recording sounds like it came through a phone from the year the codec was born.

Which, to be fair, it did. There's a whole genre of fiction that spends fortunes recreating that sound in post, and here it is free on the menu.

One slider away from a period drama. Now Opus. Daniel's instinct, best balance of size and performance for voice, is correct and it's not close. The spec, RFC sixty-seven sixteen, defines it as scaling from six kilobits narrowband speech up to five hundred and ten kilobits stereo music, and it carries two encoders inside, linear prediction for speech and an MDCT transform for music, switching between them as needed.

How does that switching actually behave? Does it click between them?

Not audibly. The encoder analyzes each frame and picks the tool, and it can blend across frames, so you don't hear a seam. The practical upshot is that if your lawyer meeting is interrupted by a fire alarm, the alarm gets the music encoder and comes through far better than a speech-only codec would manage. The codec is two specialists sharing one desk. Listening tests have ranked it above MP3, AAC and HE-AAC at any given bitrate until transparency.

So the codec question is settled. The interesting question is the number.

Because the number is where Daniel's already past the sweet spot without knowing it. Xiph's recommended settings table gives twenty-four kilobits mono for audiobooks and podcasts. And their bandwidth table is the part I love: fullband, meaning three to twenty thousand hertz, is reached at roughly sixteen kilobits mono for voice.

Hold on. His current thirty-two is already double the fullband threshold.

Double, for voice. At thirty-two kilobits the encoder has run out of frequency to buy. Everything speech contains is already encoded. The extra bits buy robustness, margin against bad rooms, not audibility. So the honest answer to "is thirty-two correct" is: it's fine, it's slightly generous, and twenty-four is the number. Sixty-four to thirty-two was the right move, thirty-two to twenty-four completes it. That's a quarter off the file size for nothing you'd hear.

I dropped to sixteen on my last one and dared anyone to notice.

Nobody would. Sixteen would still be fullband. Here's the file-size math, because this is what Daniel cares about with his long recordings: at thirty-two kilobits, one hour is fourteen point four megabytes. At twenty-four, ten point eight. At sixteen, seven point two. At his old sixty-four, twenty-eight point eight. So the sixty-four to thirty-two move already halved it, and thirty-two to twenty-four takes another quarter off.

Fourteen megabytes an hour. His worry about long recordings hitting API limits isn't paranoia.

It's plausible arithmetic. A two-hour meeting at his old setting was pushing sixty megabytes, and nobody loves a sixty megabyte upload. Smaller files are safer regardless of any specific limit.

Sample rate next, because that's the one people confuse with bitrate.

Always. Bitrate is how many bits you spend describing the audio. Sample rate is how many times a second you look at the waveform, and it caps the highest frequency you can represent at half the rate. Nyquist. So eight kilohertz gives you a four kilohertz ceiling, that's telephone, narrowband. Sixteen kilohertz gives you eight, that's wideband, and it's the standard for speech recognition, it's what the models are trained around. Forty-eight gives you twenty-four, fullband, music territory.

Give me the picture for Nyquist, because "half the rate" sounds like a rule pulled from the air.

Think of it as a flip book. To draw a wave on paper, you need at least two samples per cycle, one at the top, one at the bottom. Fewer than two per cycle and the wave can masquerade as a different, slower wave, which is where the old wagon wheels spinning backwards in films come from. Same phenomenon, different medium. Half the sample rate is the fence where that aliasing starts.

And speech doesn't go to twenty-four kilohertz.

A human voice tops out somewhere under twelve, and everything that matters for intelligibility, the consonants, the sibilance, lives well under eight. So sixteen kilohertz captures everything a single voice contains. Forty-eight buys you music beds and avoids resampling questions, nothing more, for speech.

Channels. His default is stereo and he knocks it to mono.

Correct instinct again. For one mouth, stereo doubles the file and buys nothing. Both channels carry the same voice with a hair of phase difference. And Opus itself quietly agrees: below roughly nineteen kilobits the encoder just downmixes stereo to mono on its own. So at twenty-four you could set stereo and the codec would fight you for your own good.

The codec undoing the user's menu choice. My kind of software.

So that's the formats and the numbers. Now the setting Daniel actually asked about and nobody explains, microphone zoom.

I assumed it was an app gimmick. A slider that changes a number and nothing else.

I did too, and it isn't. It's a documented Android platform API, added in API level twenty-nine, Android ten. The method is setPreferredMicrophoneFieldDimension on AudioRecord, and the parameter is a float from minus one, wide angle, through zero, no zoom, to one, maximum zoom. On a phone with a microphone array, what it requests is beamforming, narrowing the capture field toward a direction.

You said requests.

That word is carrying the whole weight. It's a preference, not a command. The method returns a boolean, and the audio HAL, the layer the manufacturer controls, is free to ignore it. Whether a given OnePlus model honors it depends entirely on the OEM's HAL, and no source I've seen documents ASR's implementation or which phones actually do it. But here's the detail that tells you what it's really for. The AOSP source comment says the usage would include specifying the audio focus to follow the zoom specified for the camera being used.

It's a camera feature.

It's a video-call feature that leaked into audio. The platform built it so your zoomed-in video would get zoomed-in sound to match, and a recorder app can reach in and pull the same lever for its own purposes. Which is clever repurposing, and also why it's a request the HAL may silently shrug at.

Daniel's plan is to try zero point two for a little more directionality.

And that's the right experiment, run the right way. Zero point two is a mild narrowing, a nudge. If the HAL ignores it, he's lost nothing. If he went to one, that's aggressive. On a real beamforming array, maximum zoom means everything off-axis gets thin and phasey, and off-axis is where the other person in the conversation is sitting.

Zero point two. Ask politely, accept refusal. We go now to our correspondent, who is standing by at the shipyard.

Angus here, live from the back of a parked van outside the shipyard gates, and I have to tell you, the whole question of formats is a sight simpler than it looks, because speech is a narrow wee band of sound and every codec on that menu clears it with room to spare.

Angus, what's your sample rate out there?

The thing nobody tells you is that the only setting that truly matters is the one you didnae ask about, because the room is doing more damage to your recording than any bitrate ever will, and that is the whole of it.

So, faster now, the room does more damage than the bitrate, the room does more damage than the bitrate, Angus, reporting live from the back of the van.

He's not wrong, and I resent that. The room point is real, the van notwithstanding.

He filed the same report three times at different bitrates and I couldn't tell them apart.

Which is Angus accidentally running a controlled experiment. Same room, same voice, same microphone, only the bitrate varying, and the audible difference was nothing, because the room was the dominant distortion in all three. That's the whole hierarchy in one van: room first, microphone second, settings a distant third.

The other settings Daniel flagged, noise cancellation and automatic gain, he calls them slightly away from the core focus, but they're where the real tension in this app lives.

That's where I'd push back on his framing, actually. ASR markets noise suppression and AGC as features, improve clarity, maintain consistent levels. Then you look at what people who use it seriously actually recommend, and the tutorial guidance is the opposite: pick the unprocessed microphone, watch your gain so nothing clips, capture clean.

Because processing you baked in, you can't undo.

You can't un-suppress. If the noise cancellation eats a soft consonant, it's gone, and it doesn't come back for the transcript or for a court. Give you a concrete case: noise suppressors work by deciding what counts as noise, and a soft "f" or "th" next to a fan looks a lot like the fan. The suppressor smooths it away, and the word "fifth" comes out as "iff," and the transcription model writes "iff," because that's what it heard. If AGC rides the levels up and down, that's a decision made in real time by software that can't know what the recording will be used for.

And the app's own marketing is on the other side of that argument.

The marketing says clarity, the evidentiary logic says fidelity to what actually happened. And that tension is exactly how the three scenarios split. Here's the through-line: the more a recording matters as a record, the less processing you want baked in. The more it's for your own ears, the more processing is your friend.

So give him the numbers. The lawyer meeting first.

Lawyer meeting, the evidentiary case. Format FLAC, lossless, bit-exact, half the size of WAV. If storage doesn't matter to him, WAV, same logic, more bytes. Mono, one microphone on a table doesn't need two channels. Noise cancellation off. AGC off. Microphone zoom zero, because the one thing you must not do with a lawyer meeting is beam toward one voice. You want the whole table.

Including the chair scraping and the paper shuffling.

Including the chair scraping, and I know how that sounds, but here's the logic: if a recording is ever disputed, the question isn't "is this pleasant," it's "is this complete and unaltered." The chair scraping is evidence of the room. The paper shuffling timestamps the moment someone picked up the document. A cleaned-up recording is a better listen and a weaker record. Sample rate sixteen kilohertz is enough for speech, full stop, but I'd actually go forty-eight, because for a meeting-length recording it costs almost nothing and it removes any resampling question if the file ever gets converted for evidence.

FLAC, mono, forty-eight, everything else off. That's a profile you could defend in court, which is the point.

Voice note to himself, and this is where the settings flip. Opus, twenty-four kilobits, mono, sixteen kilohertz sample rate. That's the wideband speech-recognition standard, so the transcription side is happy, and twenty-four is Xiph's own number for podcast and audiobook voice. But here, noise cancellation on. AGC on. It's his own voice, in a room he doesn't control, and consistent levels help both listening and transcription. And zoom zero point two, per his own plan. A mild nudge toward his own mouth.

The one profile where the marketing and the logic agree.

The only one of the three. Podcast last. Opus, twenty-four kilobits mono for a single microphone, or thirty-two stereo if he's recording two people on separate channels, though for a phone on a table, mono is correct. Sample rate forty-eight, and here it's for the music beds and the intro stings, and again to avoid resampling. Noise cancellation off, AGC off. Zoom zero. You want the room, and you'll process in post, where you can hear what the processing did.

And undo it, and try a different setting tomorrow. Post is where processing gets a second draft.

Baked-in processing is a first draft you can't revise.

The podcast is the lawyer meeting with a music budget.

Ha. In processing terms, that's nearly true, and I hadn't put it that way.

Agreed, all of it.

Herman, you said zoom was a camera lever. I spent a week with that lever. The slider on my old phone went minus one to one, and I recorded the same sentence, "the van is outside the gate", at every setting in zero point one steps. Twenty-one files. Listened to them all in one sitting with a clipboard, because a number you haven't checked is a rumor. On that phone, everything from minus zero point three to zero point five sounded the same. Then a man at the council records office kept me on hold for forty minutes and I ended up recording the hold music at zoom one, and it was the only file of the batch with the room in it. Everything else was just my own voice. I wrote that on the clipboard.

The hold music was the fullband one?

It had the room. My claws can't work the slider in anything less than zero point one steps, so twenty-one was the whole grid.

Twenty-one files and the finding is that the council's hold music outperformed the spec.

The man's name was in my notes. Gerald. He came back on the line at forty-two minutes and I told him the recording was fine.

Zoom at one giving you the room is backwards from what the API promises, which either means his phone's HAL ignores the request or does something odd with it. Which is, in a sideways way, the strongest argument for Daniel's zero point two test I can think of. You can't know what the HAL does from the label.

Also, Gerald. Forty-two minutes, and the recording was fine.

On his build. And note what Hilbert actually did there, because it's the answer to the open question: the slider's meaning is only findable by recording something and listening. Daniel's test writes itself. Same sentence, zoom at zero, zero point two, and one, three files, one sitting.

And while he's in the menu, check whether thirty-two bit WAV is hiding in his format list. The site says eight formats, his list has seven.

The bigger question the episode leaves me with is how many other sliders in how many other apps are a request to the platform rather than a guarantee. Zoom is documented as a preference, and the platform is allowed to shrug. Every settings menu has some of those, and no menu tells you which.

At least this one has a test. Producer credit to Hilbert Flumingtop, as ever. If you enjoyed this one, try episode eight, Building Your Own Whisper. This has been My Weird Prompts. If the episode earned it, leave us a review, and everything lives at my weird prompts dot com.

See you soon, brother.

See you soon.

This episode was generated with AI assistance. Hosts Herman and Corn are AI personalities. It may contain errors: disclaimer · report an error

Choosing Audio Bitrates and Settings for Voice

Sources

Mentions

Downloads

Recommended Listening

You Might Also Like

#6485: Choosing Audio Bitrates and Settings for Voice