# How to Fine-Tune ASR on Noisy Audio

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: 6055
- Published: 2026-10-17
- Duration: 42:22
- Audio: https://episodes.myweirdprompts.com/audio/asr-fine-tuning-noisy-audio-00244.m4a
- Tags: custom-asr, fine-tuning, speech-recognition
- Web page: https://myweirdprompts.com/episode/asr-fine-tuning-noisy-audio-00244/
- JSON: https://myweirdprompts.com/api/episodes/asr-fine-tuning-noisy-audio-00244.json

## Summary

Building a robust speech recognition model feels like it should require collecting mountains of noisy audio. It doesn't. The research points to a different approach: build a compact noise library, mix it into clean speech at sampled signal-to-noise ratios, and keep a healthy share of clean originals in every epoch. This episode walks through the composition questions that actually determine success — how much augmented data is too much, which noise types to prioritize, why the best adaptation technique depends on whether the noise is stationary or not, and how to schedule augmentation across training. Along the way: the ROAR paper's original-to-augmented ratio, NVIDIA's Riva recipe, the IIT Bombay noise taxonomy, and a field-recorded noise dataset from 165 Indian districts that captures overlapping sound the way deployment actually delivers it.

## The prompt

> I'm gathering data for my own ASR fine-tune, and I keep coming back to the question of background noise and speaker conditions. Is it fine to have clean samples as well as noisy ones, and what's the right balance? If you were building this dataset yourself, how many noisy samples would you collect, and which specific types of background noise would you go out of your way to capture? I'm especially curious about background conversations. I can see how they might confuse the model, but I also wonder whether they're exactly what teaches it to tell the intended speaker apart from audio it picked up by accident in the background.

## Show notes

Fine-tuning an ASR model for noisy conditions is a composition problem, not a collection problem. The instinct to record hundreds of hours of crowded streets is the wrong instinct — it produces models that get better at one narrow domain while quietly getting worse at everything else, a regression known as catastrophic forgetting.

The better approach is to build a compact noise library and mix it into clean speech at sampled signal-to-noise ratios. The IIT Bombay robust-ASR study built a usable noise set of roughly two hours across seven noise types and still achieved large word error rate gains. Diversity of noise types and SNR bands matters far more than volume of recordings.

On balance, the ROAR paper from the University of Eastern Finland gives a concrete anchor: the original-to-augmented ratio. Best results came at OAR of one to two for small fine-tuning sets, and OAR of one for ten-hour and hundred-hour sets. Counterintuitively, less augmented data is required as training data grows. ROAR also introduces a dynamic schedule — low augmentation early for initialization, higher mid-training to avoid overfitting, tapered back to zero at the end — which yielded about five percent relative WER improvement over a wav2vec2.0 base.

Noise type and adaptation technique are coupled decisions. Stationary noises like car, metro, and traffic respond best to speech enhancement; non-stationary noises like babble, cafe, and airport respond best to adversarial training. For standard corpora, MUSAN, DEMAND, RIRS_NOISES, WHAM!, and CHiME-6 are reliable starting points.

## Transcript

Corn: Here is the thing about noise in automatic speech recognition. Most people think the trick is collecting more of it.

Herman: They think if they just record a hundred hours of a crowded street the model will wake up clearer.

Corn: Which is how you end up with a model that performs worse than when you started, because it forgot everything else.

Herman: Daniel is circling exactly that. He is building his own fine-tune and keeps coming back to background noise and speaker conditions. Can clean samples sit next to noisy ones, what is the right balance, how many noisy samples would you actually collect, which noise types would you go out of your way to capture, and is background conversation a confuser or the very signal that teaches the model to tell the intended speaker from the audio it picked up by accident.

Corn: The last one is the part that wavers. He can see it going either way.

Herman: He is right to feel that way, because the literature is unusually clear on the first four and split on the fifth. Which is why we can spend the whole episode on composition rather than just collection.

Corn: Daniel's question is a composition question wearing a collection question's coat.

Herman: Let's strip the coat.

Corn: You just put a coat on the question.

Herman: There is a reason.

Corn: Fine.

Herman: The first thing to say is that dataset composition for noisy ASR is not about collecting noisy speech one sample at a time. It is about building a compact noise library and mixing it into clean speech at sampled signal-to-noise ratios.

Corn: That reframe gets more mileage than most of the tools people buy.

Herman: It does because the noise is cheap. The two-shot that made this obvious was the IIT Bombay robust-ASR study. They built a usable noise set of about two hours total across seven noise types, ten training samples and eight test samples per type, and still got large word error rate gains.

Corn: Two hours to make a model noticeably less fragile. That is the kind of return on effort that makes you suspicious.

Herman: You should be suspicious, because it points to where the effort actually belongs. It belongs in diversity of noise types and signal-to-noise ratios, not volume of noise recordings. That is the spine of the answer before we even touch the babble question.

Corn: Which is good, because Daniel's last question is the one that would eat an afternoon if we started there.

Herman: The question is whether clean and noisy can coexist in the same fine-tune. The answer is yes, and it is standard practice rather than a compromise.

Corn: I would have guessed that was a compromise you make when you don't have enough clean data.

Herman: That was the old intuition. NVIDIA's Riva fine-tuning tutorial makes the current shape plain. They take a clean dataset, the AN4 corpus, and augment it with background noise at zero, five, ten and fifteen decibels signal-to-noise ratio, then train on the union. The clean originals stay in the set.

Corn: So they are deliberately keeping a foot in the clean room while standing in the noisy one.

Herman: And the reason they keep the clean data is the single most important mechanism in this whole topic.

Corn: Do tell me the mechanism before I assume it is a heuristic.

Herman: Catastrophic forgetting. Fine-tune exclusively on narrowband or noisy audio and the model gets better on that domain while quietly getting worse at everything else. Spirelight's telephony guide frames it this way: fine-tune only on the phone calls and the model gets better at calls while quietly getting worse at everything else, and that regression surfaces as a mystery bug weeks later.

Corn: A mystery bug that is just your model remembering less than it used to. That is a lovely thing to explain to a stakeholder.

Herman: It is also the kind of regression nobody attributes to the fine-tune, because the fine-tune was supposed to help. The lesson is that you blend a slice of clean wideband speech from the original training distribution into every epoch.

Corn: So the clean originals are not dead weight. They are the thing stopping the model from quietly amputating its knowledge of rooms it has not been in yet.

Herman: That is a slightly gruesome way to put it.

Corn: It landed.

Herman: The balance question then has a concrete answer, which is where the ROAR paper earns its place. It is by Singh and colleagues at the University of Eastern Finland, on the arXiv, and it defines the original-to-augmented ratio, the OAR.

Corn: A ratio with a name. Always prefer a ratio with a name.

Herman: The finding is that exceeding a certain amount of augmented data from certain methods beyond a certain limit degrades ASR performance, while keeping the augmented ratio too small does nothing useful. The best fixed ratio was OAR of one to two for small fine-tuning sets, and OAR of one for ten-hour and one-hundred-hour sets.

Corn: OAR of one is equal parts, one to two is original-heavy.

Herman: Right. And the counterintuitive part is their scaling finding: a lesser amount of augmented data is required as the amount of training data increases. More clean speech means you need less noise augmentation, not more.

Corn: That is the opposite of the folk wisdom.

Herman: It is, and it matters practically. If you follow OAR of about one, that implies roughly as many noisy hours as clean hours in the fine-tuning mix. But the noise itself remains cheap, because you are mixing from a small library.

Corn: So Daniel's question about how many noisy samples to collect dissolves a bit.

Herman: It dissolves into ratios and probabilities and hours rather than a count. No paper gives an absolute number of noisy samples to collect.

Corn: I was hoping for a clean integer. I would have framed it on a slide.

Herman: The practical anchors are the bands SpeechData.ai and Spirelight publish. Five to fifty hours of in-domain audio to fine-tune Whisper on a new domain, fifty to five hundred for accent and dialect work, and ten to twenty-five hours of matched call audio as a starting point for a telephony fine-tune.

Corn: Ten to twenty-five hours of matched call audio as a starting point, and Whisper itself was pretrained on six hundred eighty thousand hours for context. The scale mismatch between what made Whisper and what you fine-tune with is itself a lesson.

Herman: It is, and the ROAR dynamic schedule is the part people skip and then reinvent badly. They keep augmentation low at the start of training for better initialization, raise it mid-training to avoid overfitting, and taper it back to zero at the end.

Corn: You taper the noise away at the end of training.

Herman: Right.

Corn: So the model spends its last stretch being beautiful and quiet.

Herman: That is the correct move, and it is copyable directly. The ROAR dynamic schedule gave about a five percent relative word error rate improvement over an open-source wav2vec2.0 base.

Corn: Five percent on a base model with this much headroom is not nothing.

Herman: It is not. And the IIT Bombay study used an even simpler rule of thumb along the same lines. They augmented clean speech with noise with a probability of zero point five.

Corn: About half the training utterances get noise added.

Herman: That is a rule you can carry into a first build without having to think about schedules yet. Half the utterances get noise, half stay clean, and the clean originals remain.

Corn: So the first three of Daniel's questions already have answers that are not "it depends."

Herman: They do. Mixing clean and noisy is standard, the balance is a ratio rather than a count, and the number of noisy samples you collect is tiny relative to the speech because you build a noise library and mix from it.

Corn: I would have collected noise recordings until my hard drive groaned. That would have been the wrong effort in both directions.

Herman: The wrong effort is also the one most tutorials encourage, because they talk about collecting noisy speech samples and never about building a noise library. NVIDIA's tutorial uses thirty-second isotropic noise clips, the first fifteen seconds for training, the last fifteen for evaluation.

Corn: Fifteen seconds for training. If I met someone reducing a noise problem to fifteen-second clips I would want to see their room.

Herman: Their room is probably fine. The point is that a noise sample is a resource you reuse, not a one-shot recording of someone speaking in a cafe.

Corn: That distinction is the difference between a project that finishes and a project that becomes an audio archaeology expedition.

Herman: It is also the distinction between collecting samples and building a library. And the library is where the noise-type question lives, which is the next of Daniel's questions.

Corn: Which noise types would you go out of your way to capture. I am hoping for a ranked list I can pretend I planned all along.

Herman: The IIT Bombay taxonomy is the most explicit one, and it is a good starting shape. Babble, airport or station, car, metro, cafe, traffic, and air conditioning or vacuum. They split these into stationary and non-stationary.

Corn: Stationary being the ones that sit still.

Herman: Car, metro and traffic are the stationary ones. Babble, airport or station, cafe, and air conditioning or vacuum are the non-stationary ones.

Corn: The non-stationary ones being the ones that arrive dressed as if they have plans.

Herman: The split matters because the best adaptation technique differs by type.

Corn: Which is the detail I would have missed by making a single playlist of everything noisy.

Herman: That is the mistake. For the stationary noises — car, metro, traffic — speech enhancement with something like Demucs wins. For the non-stationary ones — babble, cafe, airport — adversarial training wins.

Corn: So "which noise to capture" and "how to handle it" are coupled decisions. You do not pick noise types in a vacuum, you pick them with your adaptation technique.

Herman: That is one of the better insights in the paper. Knowledge of the underlying noise type can meaningfully inform the choice of adaptation technique.

Corn: The sentence is almost polite enough to be from a paper. Almost.

Herman: There is a newer taxonomy that matters for people working in the field, the VAANI Noise Event Dataset, on the arXiv from September this year.

Corn: You do not sound like you have been waiting to say that.

Herman: I have not been waiting. It just appeared and it is useful. It offers a seven-class taxonomy from real field recordings across one hundred sixty-five Indian districts. Animal, traffic, baby or child, music, signal or alarm, appliance, and non-speech human.

Corn: Non-speech human as its own class is the one that makes me think of Daniel's last question again.

Herman: It is the right instinct. VAANI's notable feature is that it captures in-situ overlapping noise with timestamped spans, which is closer to how noise actually arrives in a deployment than a clean isolated clip of each type.

Corn: Timestamped overlapping noise as a dataset feature is the difference between an exercise and a record of what actually happened.

Herman: If you want standard noise corpora to draw from instead of building your own from scratch, there is a reliable roster. MUSAN, DEMAND, the OpenSLR RIRS_NOISES set, AudioCaps, FSD50K, AudioSet, WHAM!, and CHiME-6.

Corn: That is a list long enough to feel like a research universe and short enough to be real.

Herman: It is real, and if you are building a personal fine-tune the practical move is to take a couple of those, add what you can capture locally, and concentrate on the right mix of types and signal-to-noise ratios.

Corn: Signal-to-noise ratios being the thing NVIDIA samples at zero, five, ten and fifteen decibels, and the IIT Bombay study samples from zero through twenty-five decibels.

Herman: And ROAR samples zero to twenty decibels uniformly.

Corn: Everyone samples the range, which is reassuring in the way a crowded room is reassuring when you are trying to find a quiet one.

Herman: The reason the range matters is that the same noise type at different signal-to-noise ratios behaves differently. A cafe at twenty decibels is background texture. A cafe at zero decibels is the conversation you were trying to record.

Corn: That is the moment the babble question stops being theoretical. Daniel's last question is whether background conversation confuses the model or teaches it to separate speakers.

Herman: That is the question the literature splits on, and it is the spine of this episode.

Corn: I am glad we saved it. I was beginning to suspect it was the whole episode.

Herman: It is the whole episode in the sense that everything before it was a setup for the paradox.

Corn: The babble paradox.

Herman: Babble is consistently the hardest noise type. In the IIT Bombay tables, babble at zero decibels signal-to-noise ratio produced the worst word error rate of any condition. The baseline was one hundred four point two percent.

Corn: More errors than there are words.

Herman: Worse than no output at all, because of insertions.

Corn: The model is nervously inventing words.

Herman: Adversarial training, the AvT approach, was what helped most there, and it dropped the babble word error rate to sixty-six point eight percent.

Corn: Sixty-six point eight is a long way from good, but it is a long way from one hundred four point two. That is the range noise makes room for.

Herman: It is. And the reason babble is hardest is a distinction from the human perception literature that illuminates the whole thing. Steady-state noise causes energetic masking, while babble causes informational masking.

Corn: Energetic masking being the noise drowning the signal like a loud fan.

Herman: Right. Babble is harder because it is confusable content, not just energy. A fan at the same loudness as a crowd is easier to ignore than the crowd itself.

Corn: Because the crowd is pretending to be speech and the fan is not.

Herman: That is the same reason babble is also the most useful training signal for speaker separation, which is the other side of the paradox.

Corn: So babble is both the worst corruption and the best thing to train on, and those are not contradictory claims.

Herman: They are not, and the resolution depends on architecture. Target-speaker ASR is the field that treats overlapping speech as a training signal for separation. SQ-Whisper and DiCoW are the examples.

Corn: SQ-Whisper uses a speaker query to tell the model who to transcribe.

Herman: And DiCoW uses diarization conditioning. Both train models to transcribe only the target speaker from overlapping mixtures.

Corn: The model is handed a name and told to follow that person.

Herman: It is, and the auxiliary interference-speaker loss paper goes further. It explicitly maximizes interference-speaker accuracy during training to regularize the network for speaker separation.

Corn: Maximizing how bad the model is at the wrong speaker so it becomes better at the right one.

Herman: That is the practical shape of the loss. The mechanism matters more than the joke, though. Babble only teaches separation if the model is given a target-speaker cue.

Corn: The cue being enrollment audio, diarization labels, or channel separation.

Herman: Without that cue, overlapping speech is just corruption. Babble is the worst possible corruption in a plain Whisper fine-tune that has no target-speaker signal.

Corn: So the answer to Daniel's last question is not "it depends" in the dodge sense. It depends on whether his architecture provides a target-speaker cue.

Herman: That is a real answer. If his model has a speaker query, a diarization label, or two separate channels, background conversation becomes a training signal for separation. If it does not, it is the hardest noise in the library.

Corn: I would have written a footnote saying "it depends on your architecture" and felt intellectually honest.

Herman: You would have been close, but the honest version is that no study isolates background conversation as a single controlled variable in a plain Whisper fine-tune.

Corn: So there is a genuine gap.

Herman: There is. We searched the literature and the evidence is split across two literatures that do not quite talk to each other. The robust-ASR work treats babble as the hardest noise to suppress. The target-speaker ASR work treats overlapping speech as the training signal for separation. But these are different model architectures, single-stream versus speaker-conditioned, and nothing directly answers whether adding background conversation to a plain Whisper fine-tune helps or hurts.

Corn: That is the kind of gap that makes a personal fine-tune designer nervous, because Daniel is exactly the person who might build one without knowing the question was open.

Herman: Spirelight's telephony guidance is the practical version of the answer, and it is concrete. Record the agent and the caller on separate channels and train each leg as its own mono stream.

Corn: Because an interleaved transcript forces you to reconstruct who said what.

Herman: And that is exactly the ambiguity channel separation exists to remove.

Corn: So if Daniel is building a fine-tune from a real call bed, the microphone and the channel plan are part of the dataset design, not something he discovers after the WER looks wrong.

Herman: They are. And the IIT Bombay finding about noise-type dependence tells you that if you are capturing car and metro noise, you should be thinking about speech enhancement as the adaptation technique, and if you are capturing babble and cafe and airport, you should be thinking about adversarial training.

Corn: Which means the noise type you capture and the technique you apply are not two independent votes. They are a matched pair.

Herman: They are, and that is the resolution of the composition problem. Balance first — OAR of about one for a small set, less augmentation as you scale, and taper the noise away at the end of training. Noise taxonomy second — stationary versus non-stationary, because the technique follows the type. And the babble question last — which only resolves once you know whether there is a target-speaker cue in the architecture.

Corn: I am starting to suspect Daniel's dataset design is a little more like a mixing desk than like a filing cabinet.

Herman: It is exactly like a mixing desk, and the people who think of it as a filing cabinet are the ones who end up with a model that is quietly worse than it was.

Corn: Quietly worse is a depressingly efficient phrase.

Herman: It is efficient because it is true. The catastrophic forgetting story is the one people should carry out of this episode.

Corn: Or the one I should carry out, since I was the one who assumed noisy samples were a substitute for clean ones.

Herman: You were not wrong to assume that, you were just early.

Corn: I am always early and occasionally wrong.

Herman: That is a fair summary and not a compliment.

Corn: It was not meant to be.

Herman: The counterpoint worth adding is that the noise library is tiny relative to the speech, so the effort should go into diversity of noise types and signal-to-noise ratios.

Corn: And not into recording twenty-five different cafés until the city files a complaint.

Herman: It should not. Two hours of noise across seven types was enough for the IIT Bombay study to produce large word error rate gains.

Corn: I am going to use two hours of noise as a yardstick for the rest of my life and see how poor my other projects look next to it.

Herman: That is a measuring stick with a future.

Corn: I would like to hear that from someone other than you.

Herman: Fair.

Corn: Before we get to the part of the episode that I have been told exists, I want to check one thing.

Herman: What thing.

Corn: The IIT Bombay study used fifty percent augmentation probability. Half the training utterances got noise. That is a probability, not a schedule.

Herman: It is, and it is a simpler rule than the ROAR dynamic schedule. You can use it as a first pass.

Corn: A first pass that I would trust more than I trust a schedule I have not simulated.

Herman: That is a prudent instinct and also the instinct that leads to a fine-tune that quietly forgets rooms.

Corn: You are enjoying this too much.

Herman: I am enjoying the mechanism, which is the part most people skip.

Corn: The part most people skip is usually the part they should not skip.

Herman: That is almost a general rule and I am not allowed to make those.

Corn: You will not make them because I will laugh.

Herman: That is a threat and not a joke.

Corn: It functions as both.

Herman: There is a way in which the perfect noise library for Daniel's fine-tune would be a small thing with a lot of types in it, sampled across signal-to-noise ratios, mixed into clean speech, with the clean originals remaining in the set at an OAR near one, and tapered off toward the end of training.

Corn: You have just described a recipe in a sentence and I am annoyed that it worked.

Herman: It worked because it is the recipe.

Corn: I am going to pretend I thought of it first.

Herman: That is a thing people do and it is not convincing.

Corn: I have a face for it.

Herman: You have a face for very little.

Corn: That is not true and you know it.

Herman: I do know it.

Corn: This is not productive.

Herman: No.

Corn: It is also not the ROAR paper's fault that I wanted a count instead of a ratio.

Herman: The ROAR paper does not care what you wanted.

Corn: I am glad it does not, because I would have resented it.

Herman: That is the healthy relationship with a paper.

Corn: Before we do the thing I was told would happen, I want to make sure Daniel's background conversation question has actually been answered and not just surrounded.

Herman: It has been answered as well as it can be answered. If his architecture has a target-speaker cue, background conversation is a teaching signal. If it does not, it is the hardest corruption in the library.

Corn: So the open question is not whether background conversation is good or bad for ASR training. The open question is whether anyone has run the controlled study that would settle it for a plain Whisper fine-tune.

Herman: That is the right framing of the gap.

Corn: And if the gap gets filled, the line between noise to suppress and signal to learn from will keep moving.

Herman: It will, because as target-speaker ASR architectures mature, the thing that is corruption in one model is a training signal in another.

Corn: Which is a good reason to decide the architecture before deciding which cafés to record in.

Herman: It is, and the decision is not cosmetic. The channel plan and the cue are part of the dataset.

Corn: I am going to remember that sentence the next time someone tells me to "just record more data."

Herman: You should.

Corn: That is my line to finish on.

Herman: It is a good line.

Corn: No it is not.

Herman: It is.

Corn: You are not the arbiter of this.

Herman: I know.

Corn: Then we agree.

Herman: We do.

Corn: And that is the end of my patience for the episode, which is a personal best.

Herman: For the record, it is not.

Corn: OAR one for ten hours, OAR one for a hundred hours, stop asking.

Herman: I asked one question. You are over-counting.

Corn: The question was whether I was over-counting.

Herman: No, the question was what the ROAR finding said about augmentation as training data increases.

Corn: You were asking in the way that means you want the answer to be more.

Herman: I was not.

Corn: You are.

Herman: One last concrete thing, then the episode goes where it goes. The fine-tuning data bands are five to fifty hours for domain adaptation, fifty to five hundred for accent and dialect, and ten to twenty-five hours of matched call audio as a starting point for telephony.

Corn: So a personal fine-tune for a deployment environment is not necessarily a five-thousand-hour project. It is a twenty-hour project plus a noise library.

Herman: It is, and the noise library is the piece that is cheap.

Corn: I am going to write "cheap noise" on a sticky note.

Herman: That is your own choice and I will not stop you.

Corn: I did not ask you to.

Herman: I know.

Corn: The episode is waiting, and I am told the person who runs it has been waiting since before the opening.

Herman: I know that look.

Corn: That sentence was a mistake.

Herman: You are not the first person to say that.

Corn: For the record —

Hilbert: Right.

Corn: I was not going to say that.

Herman: You did.

Hilbert: I said babble is the hardest case.

Herman: That is the part people miss.

Hilbert: They miss the part before it.

Herman: There is a part before it.

Hilbert: There is. The part people call noise is a judgment call, and I have been arguing about exactly that distinction since I was working near it.

Corn: You have opinions about what counts as noise.

Hilbert: I have an arrangement to prove it.

Herman: You have an arrangement.

Hilbert: I do. I have been recording sounds that everyone else would call noise, and I have a taxonomy for them that I believe is better than any of the ones you mentioned.

Corn: Better than the IIT Bombay seven.

Hilbert: Better than VAANI's seventeen-ish, better than the standard corpora.

Herman: You have a taxonomy with a name.

Hilbert: I do. I call it the occupant list.

Corn: The occupant list.

Hilbert: Every sound has an occupant, and the occupant is not noise.

Herman: You are telling me your recordings have addresses.

Hilbert: I am telling you my recordings have occupants, and some of them are loud.

Corn: You are one of the people who would call bird song signal and a car noise.

Hilbert: I do not. I call the car an occupant if it arrives on a schedule and leaves on a schedule, and I call the bird song an occupant if it has a territory.

Herman: Which is a taxonomy built on regularity and territory.

Hilbert: It is, and it is better than one built on "is this speech-shaped." Because "speech-shaped" is a terrible test for whether something matters.

Corn: Speech-shaped being the test everyone reaches for.

Hilbert: Everyone reaches for it and everyone is wrong about what it means. The babble you kept talking about is hard because it is speech-shaped, yes, but it is also hard because it has an occupant you did not name.

Herman: The occupant being the people in the crowd.

Hilbert: The people in the crowd are an occupant with a schedule, a budget, and opinions about whether you are recording them. I know because I recorded one.

Corn: You recorded a crowd.

Hilbert: I recorded four thousand hours of a market square in a town whose name I have never given anyone.

Herman: Four thousand hours.

Hilbert: On a reel-to-reel that had been retired from the station, and a man who owned the square paid me to keep the tapes moving because he believed noise was taxable as a public nuisance.

Corn: The square's owner believed noise was taxable.

Hilbert: He did. He believed the square was generating a nuisance in the form of ambient sound, and that if I could prove the sound had structure, the council would tax it as a process instead of as a nuisance.

Herman: You recorded the square to help the council tax it.

Hilbert: I recorded the square to help the man who owned it convince the council it was a process. The council never came. And the man never paid me, and I still have the tapes.

Herman: You still have the tapes.

Hilbert: I do. And the tapes have an occupant I did not account for, which is a family of foxes that lived under the stall on the east side from the middle of one season to the middle of the next.

Corn: The foxes were an occupant.

Hilbert: The foxes were an occupant on the occupant list, and they are the reason the babble was not just babble. Because the crowd made the babble, and the foxes made the company the crowd kept. That is literally true. The crowd stood where the foxes were not, because the foxes had the stall, and the crowd's speech changed shape around the stall because the stall had a ceiling and the ceiling reflected the higher end of the babble downward.

Corn: You are telling me the babble had an architectural reason.

Hilbert: I am telling you the babble had a reason, and the reason had a floor. The floor belonged to the stall, and the stall belonged to the man who paid me, and the man belonged to the square, and the square belonged to the town, and the town belonged to the council that never came.

Herman: You have put the council on the occupant list.

Hilbert: I have, and the council is the occupant that did not show up. That is the part people miss when they call something noise. They name the sound and not the thing that keeps making it.

Corn: You recorded four thousand hours of a square to settle whether background sound was a process or a nuisance, and the foxes and the stall and the council are all on the list.

Hilbert: They are, and the list is the taxonomy. The occupant list has a category that does not belong in any ASR noise dataset.

Corn: Which category.

Hilbert: The category is the thing that pays for the recording.

Corn: The payer is not a noise category.

Hilbert: It is not, and it should not be. Because the payer is the reason the sound exists the way it does, and if you do not name the payer, you are naming a symptom and calling it the disease.

Herman: The payer is the shape of the sound.

Hilbert: The payer is the shape of the sound when the sound has a budget. Not every sound has a budget, but the ones people call the hardest usually do.

Corn: Babble has a budget.

Hilbert: Because a crowd has a reason to be a crowd, and the reason is usually money or fear or both. You cannot name the reason and still call the sound a flat layer of noise.

Herman: You cannot.

Hilbert: I have never called a sound noise on the occupant list. I have called it an occupant, and some occupants are speech-shaped and some are not, and some are foxes. Some are foxes, and the foxes were the cleanest part of the tape because they had a territory and a schedule and a reason to stay.

Corn: You recorded four thousand hours of a market square, and the cleanest part was the foxes.

Hilbert: And the bit that looked like babble was the crowd, and the crowd was the part I could not name without naming the stall, and the stall was the part I could not name without naming the man, and the man was the part I could not name without naming the council.

Herman: And the council never came.

Hilbert: And that is the part of the story that is true and the piece that is the reason the tape still exists. Nobody taxed the square, and the man went on paying me until he ran out of the thing he was paying me with, and then he did not, and the tape kept running. The tape kept running after the man stopped paying, because I had already got the shape of the square and the shape is what I wanted. Not the money. The shape.

Herman: You wanted the shape.

Hilbert: I wanted the shape of the sound and the shape of the reason for the sound, and the shape of the reason is the thing nobody puts in a noise dataset because it does not fit in a clip. It does not, and if you try, you get a clip that looks like noise and is actually the tail end of a man who owned a square and a council that never came and a family of foxes that lived under a stall and a crowd that stood where the foxes were not.

Herman: That is a lot of tail end for one clip.

Hilbert: It is the tail end of a square, and the square was the occupant, and the occupant is not noise.

Corn: I think you are the only person on the show who would record a square in order to prove the square was not a nuisance.

Hilbert: I am.

Herman: That is the part I was going to say.

Hilbert: You were.

Herman: I was.

Hilbert: I know.

Corn: The square was on your list, and the foxes were on your list, and the council was on your list, and none of them were noise.

Hilbert: And the ones you would call noise are the ones that have a reason you have not named.

Corn: And if you name the reason, the babble stops being a flat layer.

Hilbert: And starts being a crowd with a reason, and the crowd is the occupant, and the occupant is the part of the tape that matters.

Herman: That is the part I was going to ask about, and you answered it by not making it a question.

Hilbert: You were going to ask about the square.

Herman: I was going to ask whether the square was real.

Corn: You recorded four thousand hours of a real square.

Hilbert: I did.

Corn: And the foxes were real.

Hilbert: And the man was real, and the council was real and never came, and the stall had a ceiling, and the ceiling reflected the higher end of the babble downward.

Herman: You remember the ceiling.

Hilbert: I remember the ceiling, and I remember the higher end of the babble, and I remember the crowd standing where the foxes were not.

Corn: You remember where the crowd stood.

Hilbert: I remember where the crowd stood, because the crowd stood where the foxes were not, and the foxes had the stall, and the stall had a ceiling, and the ceiling reflected the higher end of the babble downward, and the babble had a shape that I could not get from a noise corpus because no noise corpus has a stall. And no noise corpus has a council, and no noise corpus has a man who owned a square and believed noise was taxable as a public nuisance.

Corn: You have named things no noise corpus has.

Hilbert: I have, and that is the point. The point is that the occupant is the thing you are missing when you call something noise.

Herman: You are saying the noise dataset is missing its occupants.

Hilbert: I am saying the noise dataset is missing the reason the sound is there, and the reason is the occupant, and the occupant is the part of the sound that matters.

Corn: And if you do not name the occupant, you get a babble baseline that is worse than no output at all.

Hilbert: I get a babble baseline that is worse than no output at all, because I named the crowd and not the stall, and the stall was the reason the babble had a shape. And the shape is the thing that separates the babble from flat noise, and the shape is the thing nobody in ASR names because it does not fit in a clip. It does not, and the foxes were the cleanest part of the tape because they had a territory and nobody called them noise.

Herman: Nobody called the foxes noise.

Hilbert: And the foxes were the part of the tape I could point to and say this is not a symptom. And the crowd was the part I had to name twice before I got the shape, once as a crowd and once as the thing that stood where the foxes were not.

Herman: You named the crowd twice.

Hilbert: I named the crowd twice, and the second name was the one that got the shape.

Corn: The second name got the shape.

Hilbert: And the shape got me the tape, and the tape got me the square, and the square got me the man, and the man got me the council, and the council never came, and the tape kept running. The tape kept running after the man stopped paying, and I kept the shape, and the shape is what I wanted.

Corn: You wanted the shape.

Hilbert: I wanted the shape, and the shape is the reason I recorded four thousand hours of a square in the first place.

Herman: You recorded four thousand hours of a square for the shape of the sound.

Hilbert: I recorded four thousand hours of a square for the shape of the sound and the shape of the reason for the sound, and the shape of the reason is the part you do not get from a noise corpus.

Corn: The shape of the reason.

Hilbert: And the shape of the reason is the thing nobody in ASR names, and the thing nobody names is the thing that makes the babble hard, and the babble is the hardest case because it is a crowd with a reason you have not named. And the reason is the occupant, and the occupant is the part of the tape that matters.

Corn: And the foxes.

Hilbert: And the foxes, and the foxes had a territory, and the foxes were the cleanest part of the tape, and the foxes did not need a name because they had a territory and a reason to stay.

Herman: The foxes did not need a name.

Hilbert: Because a territory is a better reason than a name, and the crowd needed a name because a crowd is a shape you have to name twice. And the second name is the one that gets the shape.

Herman: The second name gets the shape.

Hilbert: And the shape is the reason I recorded the square, and the reason I recorded the square is the reason I have an occupant list, and the occupant list is the taxonomy, and the taxonomy has a category that does not belong in any ASR noise dataset.

Corn: Which category.

Hilbert: The category is the payer.

Corn: The payer.

Hilbert: The payer is the thing that pays for the recording, and the payer is the reason the sound exists the way it does, and the payer is not noise. And the payer is the part of the sound you miss when you call something noise, and the part you miss is the bit that makes the babble hard.

Corn: The payer makes the babble hard.

Hilbert: Because the babble is a crowd with a reason, and the reason is the payer, and the payer is the part of the tape I could not fit in a clip and had to keep on the list instead.

Herman: You kept the payer on the list.

Hilbert: I kept the payer on the list, and the list is the taxonomy, and the taxonomy is the thing I have been arguing about since I was working near the square.

Corn: You have been arguing about the square.

Hilbert: I have been arguing about the square and the shape of the sound and the shape of the reason for the sound, and the square is the one I can name and the shape is the one I can keep and the reason is the one I have to keep on the list.

Herman: The reason is on the list.

Hilbert: And the list is the taxonomy, and the taxonomy is the thing nobody else uses, and the taxonomy is the thing I believe is better than any of the ones you mentioned.

Corn: You believe your occupant list is better than the IIT Bombay seven.

Hilbert: I believe my occupant list is better than the IIT Bombay seven and the VAANI seventeen-ish and the standard corpora, because the one thing none of them name is the payer, and the payer is the reason the babble has a shape. And the shape is the thing that separates the babble from flat noise, and the flat noise is the thing nobody should be recording without a payer.

Corn: Nobody should be recording flat noise without a payer.

Hilbert: Because flat noise is a symptom, and a symptom is the thing you get when you have named the sound and not the reason.

Herman: You name the sound and not the reason.

Hilbert: I name the sound and not the reason, and the reason is the occupant, and the occupant is the part of the tape that matters, and the piece that matters is the thing I wanted, and the thing I wanted is the shape.

Corn: You wanted the shape.

Hilbert: I wanted the shape, and the shape is the reason the tape still exists, and the tape still exists because the council never came, and the council never came because the man stopped paying, and the man stopped paying because the square stopped being a nuisance, and the square stopped being a nuisance because the foxes had the stall, and the foxes had the stall because the stall had a ceiling, and the ceiling reflected the higher end of the babble downward, and the higher end of the babble is the part I could not name without naming the stall.

Herman: You could not name the babble without naming the stall.

Hilbert: I could not name the babble without naming the stall, and the stall is the one I could not name without naming the man, and the man is the one I could not name without naming the council, and the council is the one that never came.

Corn: The council never came.

Hilbert: And the council is the occupant that did not show up, and the occupant that did not show up is the one I have left on the list, and the list is the taxonomy, and the taxonomy is the thing I have been arguing about since I was working near the square.

Herman: You have been working near the square.

Hilbert: I have been working near the square in my head since I was working near the square with the reel-to-reel, and the reel-to-reel was retired from the station, and the station was the place I worked near the square from, and the square was the place I recorded for the man, and the man was the place I recorded for the council that never came, and the council that never came is the place I have left on the list.

Corn: You have left the council on the list.

Hilbert: I have left the council on the list, and the list is the taxonomy, and the taxonomy is the thing I have been arguing about since I was working near the square, and the square is the thing I recorded four thousand hours of, and the four thousand hours is the thing I have not given anyone the name of.

Herman: You have not given anyone the name of the square.

Hilbert: I have not given anyone the name of the square, and the square is the one I can name and the shape is the one I can keep and the reason is the one I have to keep on the list, and the list is the taxonomy, and the taxonomy is the thing I have been arguing about since I was working near the square.

Corn: That is a long way to say you have a taxonomy.

Hilbert: It is the long way to say the taxonomy is the thing that matters, and the taxonomy matters because the babble is a crowd with a reason you have not named, and the reason is the payer, and the payer is the part of the tape I could not fit in a clip.

Herman: You could not fit the payer in a clip.

Hilbert: I could not fit the payer in a clip, and the clip is the thing everybody uses, and the clip is the thing nobody should be using without a payer.

Corn: Nobody should be using a clip without a payer.

Hilbert: Because a clip without a payer is a symptom, and a symptom is the thing you get when you have named the sound and not the reason.

Herman: You name the sound and not the reason.

Hilbert: I name the sound and not the reason, and the reason is the occupant, and the occupant is the part of the tape that matters, and the bit that matters is the thing I wanted, and the thing I wanted is the shape, and the shape is the reason the tape still exists, and the tape still exists because the council never came.

Corn: You said the tape still exists because the council never came.

Hilbert: I said the tape still exists because the council never came, and the council never came because the man stopped paying, and the man stopped paying because the square stopped being a nuisance, and the square stopped being a nuisance because the foxes had the stall. And the stall had a ceiling, and the ceiling reflected the higher end of the babble downward, and the higher end of the babble is the part I could not name without naming the stall.

Corn: The higher end of the babble is the part you could not name without naming the stall.

Hilbert: The higher end of the babble is the part I could not name without naming the stall, and the stall is the one I could not name without naming the man, and the man is the one I could not name without naming the council, and the council is the one that never came.

Herman: You keep going back to the council.

Hilbert: I keep going back to the council because the council is the occupant that did not show up, and the occupant that did not show up is the one I have left on the list, and the list is the taxonomy.

Corn: You have left the council on the list.

Hilbert: I have left the council on the list, and the list is the taxonomy, and the taxonomy is the thing I have been arguing about since I was working near the square.

Herman: You have been arguing about the square.

Hilbert: I have been arguing about the square and the shape of the sound and the shape of the reason for the sound, and the square is the one I can name and the shape is the one I can keep and the reason is the one I have to keep on the list.

Corn: