# Why Synthetic Speech Needs Fake Room Noise

My Weird Prompts — an AI-generated podcast. Daniel Rosehill sends in a prompt; AI hosts Corn and Herman discuss it. Verify important claims independently.

- Episode: 6230
- Published: 2026-10-23
- Duration: 20:38
- Audio: https://episodes.myweirdprompts.com/audio/synthetic-speech-room-tone-beds-79224.m4a
- Tags: speech-recognition, hvac, audio-equipment, text-to-speech
- Web page: https://myweirdprompts.com/episode/synthetic-speech-room-tone-beds-79224/
- JSON: https://myweirdprompts.com/api/episodes/synthetic-speech-room-tone-beds-79224.json

## Summary

Synthetic speech can stop completely between sentences, leaving gaps of true digital black that sound unnatural on playback. This episode digs into why engineers fill those holes with room-tone beds, how a real fix for this show landed around minus sixty dBFS on a phone, and why a constant quiet background beats a perfectly clean pause. Along the way: what room tone actually contains, why filmmakers record thirty to sixty seconds of nobody doing anything, and why audiobook noise-floor specs collide with loudness normalization.

## The prompt

> Why do the silences in synthetic speech need fake noise? A real recording is never truly silent: there's always room tone, mic self-noise, a little air. Text-to-speech output, on the other hand, can drop to near digital black between sentences. Audio engineers have long added room tone under edits so the gaps don't sound dead, and productions built from TTS face the same problem in a sharper form. How should you design an artificial noise floor for synthetic voices — how loud, what spectrum, recorded room tone versus generated noise — and why does dead silence sound worse on phones and earbuds than on studio monitors?

## Show notes

Real recordings always carry a room — the fridge cycling, traffic through the window, the microphone's own self-noise. Synthetic voices can stop completely between sentences, producing stretches of digital black where no room exists at all. That contrast is audible even when neither state sounds consciously noisy, which is why engineers have been filling gaps under dialogue for decades. The trick is a continuous room-tone bed: steady, unidentifiable background energy that listeners can disregard, rather than a background that switches on and off with every breath.

A fix for this show put roughly a hundred gaps per episode at minus seventy-five dBFS; listening tests on a phone preferred minus sixty, and that's where the pipeline now sits. But that's a result for this material and device, not a universal setting. The zeropointnine audiobook tool offers beds at minus seventy-five, minus seventy, and minus sixty-five dBFS measured as noise RMS before loudness normalization — and if normalization raises a quiet raw voice, it raises the bed mixed into it too, pushing the final floor past ACX's minus sixty dBFS limit.

Measurement is its own trap. In one project with five hundred generated segments, the quietest slice of a typical pause sat near minus ninety-one dBFS, while levels across the whole pause swung by about thirty decibels. Adaptive topping-up and borrowing room tone from neighboring segments both got removed for following breaths instead of a genuine bed. A constant track wins because it knows less. Start with the speech level, add a very quiet continuous bed, compare against no bed at matched volume, and measure the delivered file.

## Sources

- [zeropointnine TTS audiobook tool docs, last updated 2026-10-12](https://zeropointnine.github.io/tts-audiobook-tool/docs-dev/room-tone.md)
- [`erm` tool docs](https://doug.sh/docs/erm/denoise-and-room-tone/)
- [Descript Help Center](https://help.descript.com/effects-animations-transitions/room-tone)
- [Anand Butani, Audio Engineering × AI Experiment, 2026-04-12](https://medium.com/@anand.butani/audio-engineering-ai-experiment-3ca2fd6568e0)
- [Tomevox glossary, updated 2026-08-06](https://tomevox.com/glossary-noise-floor)
- [Jeff Talman, designingsound.org, 2014-06-05](https://designingsound.org/2014/06/05/the-negative-space-of-sound/)
- [iZotope, Basics of Room Tone](https://www.izotope.com/community/blog/basics-of-room-tone-audio-editing)
- [ZOOOP glossary, Room Tone](https://zooop.ai/glossary/editing/room-tone)
- [How the Brain Invents Silence, 2023-05-02, updated 2026-06-29](https://www.soundgearx.com/post/detail/31/)
- [DAIEN-TTS, revised 2026-01-19, ICASSP 2026](https://arxiv.org/abs/2509.14684v2)
- [extended DAIEN-TTS, 2026-08-04](https://arxiv.org/abs/2608.03011v1)
- [Filmmakers Academy glossary](https://www.filmmakersacademy.com/glossary/buzz-track/)

## Transcript

Corn: Daniel heard little fade-ins at the start of our turns, mostly on his phone. The strange part was that the fades weren't in the audio file. That led him into today's question. Why does synthetic speech sometimes need fake background noise to make the silence sound right?
Herman: The gaps were much quieter than intended. The room-tone bed was there, technically, but low enough that the playback system could behave almost as though nothing was playing.
Corn: A background employee with no duties.
Herman: Roughly a hundred gaps per episode were sitting around minus seventy-five decibels relative to digital full scale. Daniel tried louder beds on his phone, preferred minus sixty, and that's where the pipeline now puts those gaps.
Corn: His question widens out from that fix. Real recordings carry a room, microphone noise, a little air. Synthetic voices can stop completely between sentences. Engineers have been filling holes under dialogue for decades, but when there was never a room, somebody has to decide what goes there. I like that we've reached the point where silence requires a production budget.
Herman: The important distinction is between hearing steady quiet and hearing the background switch off. Those can have very different effects, even when neither sounds consciously noisy.
Corn: Particularly if the next word arrives sounding as though it needs permission.
Herman: A digital sample can be zero. A whole stretch of samples can be zero. That's digital black. A microphone pointed into a quiet booth still produces something, from the room and its own electronics. Synthetic speech can jump between those conditions, giving you a faint texture while the voice speaks and nothing underneath the pause.
Corn: My instinct would have been to clean the voice harder. If the pause is perfect, remove whatever makes the speaking sound imperfect.
Herman: That can help, but the last stretch gets expensive in damaged consonants and unnatural decays. There's an audio tool called erm that treats denoising and room tone as partners. Remove the uneven, distracting floor, then add a uniform one back.
Corn: We have purified the water and returned one ceremonial speck.
Herman: The uniformity matters. A constant, quiet background becomes something you can disregard. A background that appears at every breath keeps announcing changes. Corn, would you rather hear a slightly audible room throughout, or perfect gaps with a faint hiss only when I speak?
Corn: The room throughout. Otherwise your voice sounds like it's opening a valve. Although I'd want to check whether I'm hearing hiss, breath, or an overenthusiastic compressor before adding anything.
Herman: Those need different fixes. Breath belongs to the performance. A compressor may be pulling unwanted noise forward. A room-tone bed only helps the continuity underneath them.
Corn: There is also a limit to how much room I want. If you add a kitchen, I start waiting for somebody to unload the dishwasher.
Herman: Usually we want a texture with no identifiable event. You shouldn't be following it. Good noise-cancelling headphones make the switching especially exposed because less outside sound is available to cover it.
Corn: Your headset is creating an excellent listening environment for the fault.
Herman: Yes. Well... an excellent environment for noticing it. Whether it's objectionable depends on the voice, the bed, the listener, and the actual playback chain. Daniel's phone result gives us a practical result for this show, not a universal setting.
Corn: A real kitchen is useful here because nobody would call it noisy until they tried recording in it. The refrigerator cycles, traffic gets through the window, the building has pipes. Then you stop talking and discover the microphone has been attending a different conversation.
Herman: Room tone contains those sources, shaped by the space. Materials absorb different frequencies. Dimensions affect resonances. Air moving past an opening sounds different from air moving through a duct. Even without an obvious appliance, the room has a particular low-level sound.
Corn: You were about to diagnose our cupboards.
Herman: Only the reflective ones.
Corn: I knew we'd lose the mugs.
Herman: They can stay. A microphone adds self-noise, and the preamp adds noise too. Those become part of what remains when narration stops. Capturing a clean signal gives you more room to work later.
Corn: People sometimes say the floor is fixed at capture and can't be lowered afterward. That's a little strong. Denoising exists. It just doesn't retrospectively give you a better recording.
Herman: You can reduce the apparent floor afterward. You can't recover every quiet detail that was buried in it, and aggressive processing can leave its own sound. Better electronics don't silence a fridge, either. You have to identify which part of the chain is responsible.
Corn: The premium preamp faithfully preserves your inexpensive refrigerator.
Herman: Filmmakers commonly capture thirty to sixty seconds of room tone for a setup. Same microphone, same position, same settings, everyone holding still. Sometimes it's called a buzz track. You want the space as it sounded during that dialogue, not a vaguely similar empty room later.
Corn: Keeping people there matters too. Their bodies and clothes affect the acoustics. Empty space doesn't necessarily mean evacuate everybody and record a more reverberant room.
Herman: Silence from the people, rather than their disappearance. Moving the microphone can change the balance enough to make the patch noticeable. Changing gain changes the electronic noise along with everything else.
Corn: A surprisingly demanding recording of nobody doing anything.
Herman: Then the editor uses that sound underneath cuts, or fills a gap left by removing something. Suppose you take out a chair squeak between two lines. Replacing it with digital black can make the whole location vanish for a moment.
Corn: Even if I don't consciously hear the location, I hear its absence. That seems to explain why heavy gating bothers me. Every sentence gets its own little opening ceremony.
Herman: A gate can be useful, but an abrupt one draws a hard boundary around the voice. The background drops when the voice falls below a threshold. Expansion can make that gentler, and room tone can preserve continuity through edited pauses.
Corn: Would you put the bed only in the gaps?
Herman: Usually I'd want continuity underneath the dialogue too, at a suitably low level. Otherwise we're making another switch, just a quieter one. Crossfades smooth joins. A continuous track avoids having to reinvent the background at every sentence.
Corn: The loop still has to be boring enough to repeat. One distant bird becomes a very punctual bird over an audiobook.
Herman: Short recordings are vulnerable to that. A tiny scrape, a tonal wobble, even a distinctive fluctuation can become recognisable. You may need a different section, longer material, or carefully blended repetitions.
Corn: This isn't unique to speech synthesis. Early CDs made absolute quiet between tracks newly conspicuous compared with vinyl's continuous surface sound. Some engineers used low-level ambience to avoid that sense of the recording dropping away.
Herman: The clean medium exposed an aesthetic choice that surface noise had previously made for you. Dither is another case where deliberately adding noise improves a digital result, although it solves a different problem.
Corn: You had the dither explanation loaded.
Herman: I still have it loaded. Quantization can produce correlated distortion at very low levels. Dither decorrelates that error, making quiet fades behave better. It doesn't supply a room, and normal dither levels aren't a substitute for this bed.
Corn: You may keep that paragraph. Nothing with equations.
Herman: Generous terms.
Herman: Synthetic speech sharpens the editing problem because there may be no stable background to begin with. The model produces breaths, noisy consonants and decaying sounds, then drops almost to nothing. All of those appear in a waveform measurement, but they aren't interchangeable.
Corn: A breath is a person doing something. Treat it as the room and you're trying to keep the person permanently exhaling.
Herman: In one project with five hundred generated segments, a typical pause's quietest twenty-five milliseconds sat around minus ninety-one dBFS. That's close enough to nothing for the practical contrast to be huge.
Corn: Yet measuring a broader slice of the pause would catch the breath and give you a much louder answer.
Herman: Levels within those pauses swung by roughly thirty decibels. A pause wasn't one stable thing you could measure once and reproduce. That's why a quiet background estimate can be badly misleading.
Corn: I can already see you reaching for an adaptive system.
Herman: I would have tried one.
Corn: Herman, you would have given the silence a dashboard.
Herman: Possibly a small one. The audiobook tool we're discussing actually tried both adaptive topping-up and borrowing room tone from neighbouring segments. Both approaches were removed. Measurement mostly followed the breaths instead of a genuine bed.
Corn: A constant track wins because it knows less.
Herman: It has a simpler job. Supply steady background energy without pretending every noise in the voice tells us what the background should be.
Corn: Then my first practical decision would be how loud the finished voice is going to be. Choosing a bed before that feels like choosing the background against a moving target.
Herman: That catches an important trap. The zeropointnine audiobook tool offers beds at minus seventy-five, minus seventy, and minus sixty-five dBFS, measured as noise RMS before loudness normalization. They're engineering starting points, not perceptual thresholds established across voices.
Corn: Three levels of professionally supplied nothing.
Herman: If normalization raises a quiet raw voice, it raises a bed already mixed into that voice too. The preset label no longer tells you the final noise floor.
Corn: ACX asks for a floor at or below minus sixty dBFS. If the high preset begins at minus sixty-five, you haven't left much headroom for turning the entire mix up.
Herman: Correct. Well... headroom in the sense of margin below that noise limit, rather than peak headroom. The tool warns about that collision. Measure the delivered file, after the processing that can change it.
Corn: Our minus sixty gap setting is therefore not an audiobook recipe. Daniel was solving an audible problem on his phone, not certifying a master for every distributor.
Herman: His comparison included minus fifty-four, which he didn't prefer. Louder wasn't automatically better. Our chosen setting is a result from that listening test, with this material and that device.
Corn: What would you actually begin with if I handed you a fresh synthetic voice?
Herman: I'd establish the speech level, add a very quiet continuous bed, then compare it with no bed at matched playback volume. Increase it only until the objectionable switching stops. Check pauses and speech, and check the final export.
Corn: I would also keep a version with the bed turned off. Otherwise the first improvement becomes the standard and you start polishing something that didn't need polishing.
Herman: Would you use a recording or generated noise?
Corn: Generated for a voice with no consistent location. Recorded if we're trying to preserve an existing production. I don't want a random room assigned to my voice because its sample happened to be convenient.
Herman: That's where I land too. The erm approach harvests room tone from the original recording, not the denoised version, preferably before the first word. It trims the edges to avoid transitions and wants a clean window long enough to use.
Corn: Its default attenuation is minus twelve decibels. That's relative to the captured sample, though, not a final floor of minus twelve dBFS. Those numbers mean completely different things.
Herman: Turning down an unknown input by a known amount still gives you an unknown final level. Their suggested attenuation range extends to minus twenty decibels. You have to listen and measure in context.
Corn: Otherwise somebody will produce the loudest empty booth in publishing.
Herman: For generated beds, a useful starting spectrum is pinkish noise with the extremes controlled. White noise puts equal energy per unit of frequency, which can make the top end conspicuous. Pink noise shifts the balance downward.
Corn: I'd predict that more bass makes it feel like ventilation rather than hiss. Then I worry we've installed a fictional air conditioner.
Herman: The low end needs restraint as well. This implementation high-passes at eighty hertz and low-passes at eight kilohertz. That removes rumble below and excessive airy noise above the useful voice texture.
Corn: Useful voice texture varies by model, surely. A bright voice might tolerate a different bed from a muffled one.
Herman: The eight-kilohertz choice belongs to that implementation, not a law of synthetic speech. Listen for a separate layer of fizz that continues after the voice ends. If you hear that, the spectrum or level may be wrong.
Corn: Consistency also means not starting the generator again at every turn. A tiny repeated noise pattern could become its own audible punctuation.
Herman: They use a fixed seed for reproducibility and carry filter state between calls. Cutting the job into segments doesn't mean restarting the filter. A continuous noise stream avoids seams introduced by the noise generator itself.
Corn: Now, the phone. I think we have two different complaints being bundled together. One is that silence feels unnaturally empty. Daniel's was that the next word seemed to fade in.
Herman: Different observations, and potentially different causes. With noise-cancelling headphones, reduced outside noise makes the track's own changes more exposed. Speakers in a room have ambient sound around them, even while their input is silent.
Corn: My room is supplying the bed for free.
Herman: Your room may also mask the first tiny part of an onset. For the phone result, a plausible explanation is playback processing that changes gain or power state during very quiet passages, then takes a moment to settle when speech returns. We haven't confirmed that mechanism on the device.
Corn: Keeping a small signal alive could prevent that behaviour without changing any fade in the file. Which fits what Daniel heard, but doesn't identify the particular processor doing it.
Herman: Nor does it prove all phones or Bluetooth earbuds do it. There's no direct comparison here establishing that phones universally handle silence worse than studio monitors. Device processing, isolation and the surrounding room can each change the result.
Corn: The brain contributes without needing a mystical dislike of zeros. We're sensitive to sound boundaries. A background disappearing is an event, even when we weren't paying attention to the background.
Herman: Auditory processing remains active through gaps, and silence can be represented as a perceptual event. That supports taking transitions seriously. It doesn't hand us a scientifically ordained noise level.
Corn: Minus sixty, as revealed by the auditory cortex.
Herman: No such revelation. Corn, if you were choosing the actual room sound, would you take a booth or something more domestic?
Hilbert: ...the shop, provided Nissim stops charging me for the refrigerator. That's the latest disagreement. I've booked forty seconds of his repair shop for Tuesday. Forty seconds with the fridge unplugged, nobody moving, same microphone position. He's put refrigeration loss on the estimate. The door will remain shut.
Corn: He's charging you for the cold escaping electrically.
Hilbert: He says the fridge is part of the premises. I say I ordered the room. Herman, you separated breath from background very carefully earlier. I have separated the fridge on the booking form. Nissim has written it back in underneath. We've dealt with each other too long to keep using the same sheet.
Herman: Switching it off immediately before recording could leave clicks or settling noises. You'll need time before your forty seconds.
Hilbert: I've allowed six minutes. Tuesday I am taking the shop's silence to a second shop to have it checked. Nissim has agreed to come inside the recording, standing behind the counter throughout, because he won't accept a recording of his premises made without supervision. The second man closes at one.
Corn: Your witness is part of the acoustic treatment.
Hilbert: He has a good thick coat. I'll ask him to keep it on. I wanted a field recorder from him this week. Nine hundred and forty shekels. Didn't buy it. The little controls don't suit my claws, and he wanted extra for the case. I will use the recorder on my phone. He has already cleared me a space on the counter beside the receipt spike.
Herman: Leave enough distance that the phone isn't picking up every movement against the counter.
Hilbert: I've marked the spot on the booking form. I put his chair there too. His knee bothers him when he stands still. He crossed out refrigeration loss yesterday and replaced it with six minutes of delayed stock cooling. Forty-eight shekels.
Corn: Six minutes of waiting before forty seconds of material. That part is rather recognisable. The useful recording can be shorter than the preparation needed to keep it uneventful.
Herman: Synthetic beds avoid that particular preparation, but they still need quality control. Repeatability is valuable when you're generating a whole programme. You want to know the background won't change because a later batch used a different starting state.
Corn: There's an adjacent direction worth keeping. Environment-aware speech models try to generate the acoustic surroundings along with the voice, rather than make us glue everything together afterward.
Herman: DAIEN-TTS separates speech, background noise and reverberation, with distinct speaker and environment prompts. Its extended version also adapts signal-to-noise ratio. That could give you more coherent surroundings, though a simple steady bed remains easier to inspect.
Corn: I'd like the future version to preserve that simplicity of control. Let me change the voice without accidentally ordering a different building. For now, a quiet constant bed, judged in the finished mix on actual playback devices, is a sensible starting point.
Herman: Especially when forty seconds of usable quiet can require six minutes of waiting. Thanks to Hilbert Flumingtop for producing the show.
Corn: If you enjoyed this one, try episode fifteen, AI Gets Personal; episode four, If Your Voice Ages, Does Your Fine-Tune Become Useless; and episode ten, How ASR Went From Frustration To ... Whisper Magic. This has been My Weird Prompts, the human-AI collaboration podcast.
Herman: If you enjoy the show, leave us a review wherever you listen. It helps other curious people find us.
Corn: See you tomorrow.