Noise Floor

The lab was cold when Mara arrived at seven-fifteen the next morning, the heating system not yet convinced that anyone would be foolish enough to work this early. She kept her coat on while she booted the workstation, watching her breath make small ghosts in the air above the keyboard. The building was quiet in the way that university buildings are quiet before the institutional day begins—not silent, but emptied of intent, the background hum of ventilation and server rooms and the distant clank of pipes forming a kind of acoustic underlayment that she had long ago learned to subtract from her consciousness.

She unlocked the drawer.

The hard drive was where she had left it, of course. Data did not migrate in the night. Files did not rearrange themselves out of spite or longing. The zeros and ones sat in their magnetic domains with the patience of geology, waiting to be read, indifferent to the weight she had placed upon them.

Mara plugged the drive into her workstation and opened the session files from the previous night. Two recordings. Forty-seven seconds and thirty-nine seconds. She had listened to them raw, once each, in the field, and then she had locked them away and gone home and not slept and come back.

Now she would do the thing she actually knew how to do.

She opened her analysis software—a custom suite she had built over seven years, layer by layer, tool by tool, each module designed for a specific type of acoustic investigation. The interface was not beautiful. It was not meant to be beautiful. It was meant to be precise, and it was, with the cold precision of a surgical instrument that had been sharpened so many times it had become something other than what it was originally designed to be.

First: spectral analysis. She loaded the forty-seven-second recording and generated a spectrogram, the visual representation of frequency over time that was the bread and butter of her field. The image appeared on screen in the false-colour palette she preferred—deep blue for silence, climbing through green and yellow to a hot white for the loudest components. Most of the spectrogram was blue. The room had been quiet. The environmental noise floor sat as a thin band of green in the low frequencies, the building's HVAC system contributing its usual drone, the traffic outside adding a stochastic wash of broadband noise that peaked and ebbed with the rhythm of the street.

And there, in the middle of the recording, the anomaly.

It appeared as a cluster of bright points in the mid-frequency range—roughly 200 to 3,400 hertz, which was, not coincidentally, the frequency band that contained the fundamental and first several harmonics of the human voice. The cluster lasted approximately four seconds. It was not continuous; it pulsed, brightened, faded, brightened again, like a signal struggling to maintain coherence. Like someone trying to speak through a bad connection.

Mara stared at it for a long time.

Then she did the same thing with the thirty-nine-second recording. The second spectrogram showed a similar cluster, shorter—perhaps two and a half seconds—but in the same frequency range, with the same pulsing quality, the same sense of a signal fighting to exist against a background that did not want to carry it.

She saved both spectrograms and opened a new analysis window.

Formant extraction. This was the heart of it, the thing that had kept her awake, the thing that had driven her back to the lab at an hour when sensible people were still negotiating with their alarm clocks. Formants were the resonant frequencies of the vocal tract—the peaks in the spectral envelope that gave vowels their characteristic sound, that distinguished an 'ah' from an 'ee' from an 'oo.' They were as individual as fingerprints, shaped by the unique geometry of a person's throat and mouth and nasal passages. Change the formants and you changed the identity of the speaker. They were, in the most literal sense, the acoustic signature of a human being.

Mara had built her formant extraction algorithm during her doctoral work, refining it through three postdocs and a dozen published papers until it could pull stable formant values from recordings that would make most speech scientists weep. Recordings made in noisy environments, recordings degraded by distance and reflection, recordings that were more room than voice. She had extracted intelligible formant data from a recording made inside a running diesel engine, once, for a forensic case. She had published a paper on formant stability in reverberant environments that had been cited four hundred and twelve times.

She knew what she was doing.

She ran the algorithm on the anomalous sections of both recordings.

The first recording yielded formant values that were noisy but coherent. F1: approximately 730 hertz. F2: approximately 1,090 hertz. F3: approximately 2,440 hertz. The values fluctuated more than they would in a clean recording—the standard deviations were large, the confidence intervals wide—but the central tendencies were clear enough to work with.

The second recording was worse. The signal was weaker, the extraction less stable, the confidence intervals wider. But the central values were similar. F1: approximately 710 hertz. F2: approximately 1,120 hertz. F3: approximately 2,480 hertz.

Mara wrote the numbers in her notebook, then sat back and looked at them.

They were consistent with a female speaker. That much was unremarkable—the room had historically been associated with reports of a female voice, which was one of the reasons she had selected it for investigation. The formant values suggested a vocal tract length in the range of 14 to 15 centimeters, which was typical for an adult woman.

But formant values alone were not enough to identify a speaker. They could narrow the field, exclude candidates, provide supporting evidence. They could not, on their own, point to a single human throat and say: this one. For that, you needed comparison data.

Mara had comparison data.

She had never intended to use it. She had never intended to need it. When she had begun this project eighteen months ago—a systematic investigation of reported acoustic anomalies in buildings with specific architectural features—she had been scrupulous about maintaining emotional distance. She had chosen sites based on acoustic properties, not personal connection. She had developed protocols designed to eliminate observer bias. She had been, in every way she knew how to be, a scientist first and a grieving sister not at all.

But she had brought the comparison data with her. Lily's voice. Recordings from phone calls, voicemails, home videos. Hours of material, digitized and stored on a separate drive that she kept in her desk at home, a drive that was not part of any research protocol and served no scientific purpose and existed only because deleting her dead sister's voice was a thing she could not do.

She had brought a subset of those recordings to the lab three months ago, telling herself it was for calibration purposes, telling herself that having a known female voice sample available was standard practice, telling herself that the fact the known sample happened to be her sister was a coincidence of convenience and nothing more.

The drive was in the second drawer. She had not locked it. There had been no reason to lock it, because it was not part of the investigation, because she was not going to use it, because using it would be a breach of every methodological principle she had spent her career building.

Mara opened the second drawer and took out the drive.

She sat with it in her hand for perhaps two minutes, feeling its weight, which was negligible, and its significance, which was not. Then she plugged it in.

Lily's voice filled the lab.

Not literally. She had headphones on, and the playback volume was calibrated to a specific level, and the recording she had selected was a voicemail from three years ago—Lily calling to confirm dinner plans, her voice bright and casual and alive in the particular way that voices are alive when they do not know they are being preserved, when they are just talking, just existing, just doing the unremarkable work of being a person in the world.

"Hey, it's me. So Thursday works, but can we do seven-thirty instead of seven? I've got a thing at the gallery that runs until six-thirty and I need time to wash the paint out of my hair. Also, bring wine. Not the cheap stuff. I'm serious, Mara, if you bring that awful Sauvignon Blanc again I'm serving it to the cat. Call me back. Love you."

Mara played it through once, then again, then a third time, not because she needed to hear it three times for analysis but because it was Lily's voice and it was asking her to bring wine and it was calling her by name and it was saying love you at the end the way Lily always said love you at the end, casually, like a fact so obvious it barely needed stating.

Then she stopped the playback and ran the formant extraction.

Lily's formants, from a clean recording made on a modern phone with good signal-to-noise ratio: F1: 720 hertz. F2: 1,100 hertz. F3: 2,450 hertz.

Mara looked at the numbers.

She looked at them for a very long time.

Then she opened a statistical comparison tool and ran a formal test. The Mahalanobis distance between the anomaly formants and Lily's known formants, accounting for the higher variance in the anomaly data, was 0.73 for the first recording and 0.91 for the second. A Mahalanobis distance below 1.0 was generally considered to indicate samples from the same population. In speaker identification, distances below 1.5 were typically treated as suggestive of a match, and distances below 1.0 as strong evidence.

Both recordings fell within the match threshold.

Mara closed the comparison tool. She removed her headphones. She sat in the cold lab with her hands flat on the desk and her eyes closed and her heart doing something that she could not have described even if she had wanted to, which she did not, because describing it would have meant admitting it was happening, and admitting it was happening would have meant admitting what the data was telling her, and admitting what the data was telling her would have meant that the careful, methodical, rigorously scientific investigation of acoustic anomalies in architecturally interesting buildings had just become something else entirely.

Something that no amount of methodology could contain.

She stood up and walked to the window. The campus was beginning to wake—students moving between buildings, a maintenance cart trundling along the path below, the cafe across the street raising its shutters with a metallic clatter that she could hear faintly through the glass. Normal morning. Normal world. Normal everything except the numbers on her screen and the voice on the recording and the forty-seven seconds and thirty-nine seconds of shaped air that matched, within statistical tolerance, the acoustic signature of a woman who had been dead for two years.

Mara pressed her forehead against the glass and thought about what she was going to do.

The scientific answer was obvious. More data. The two recordings she had were suggestive but insufficient—the signal-to-noise ratio was too low, the formant extraction too unstable, the comparison too easily challenged. Any peer reviewer worth their salary would tear the analysis apart: the sample size was laughable, the environmental controls incomplete, the comparison data drawn from a source with obvious emotional bias. The fact that the comparison voice belonged to the investigator's dead sister would not just raise eyebrows; it would bury the paper before it was read.

She needed more recordings. Better recordings. She needed to deploy a proper sensor array in the Fitzrovia room—not the portable kit she had used for the initial survey, but the full system: eight microphones in a calibrated spherical array, environmental sensors logging temperature and humidity and barometric pressure and vibration, electromagnetic monitoring to rule out radio frequency interference, and a video system to document the physical state of the room during recording sessions.

She needed to record over multiple sessions, at different times of day, in different weather conditions, with different configurations. She needed to build a dataset large enough to perform robust statistical analysis, to establish whether the anomaly was reproducible, to characterize its spectral properties with sufficient precision to make the formant comparison meaningful.

And she needed to do all of this without telling anyone why.

Because the moment she said the words—the moment she told her department head or her collaborators or anyone with the power to influence her career that she believed she had recorded her dead sister's voice in an empty room—the investigation would be over. Not because the data was wrong, but because the framing was wrong, because grief and science occupied different territories in the institutional mind and the border between them was policed with the zealous efficiency of a nation that feared contamination.

So she would not say it. She would not think it, if she could manage that particular feat of cognitive discipline, which she doubted. But she would not say it. She would frame the investigation in the language of architectural acoustics and environmental monitoring and anomalous signal characterisation, and she would pursue it with the rigor and dispassion that those framings demanded, and if anyone asked why she was spending her research time recording empty rooms in Fitzrovia, she would talk about standing waves and resonant frequencies and previously undocumented acoustic phenomena, and she would be telling the truth, because all of those things were true, and the fact that they were not the whole truth was a detail she could manage.

Mara returned to her desk and began planning.

She pulled up the floor plans of the Fitzrovia building that she had obtained from the planning office. The room was on the third floor, north-facing, approximately four meters by five meters with a ceiling height of three point two meters. The walls were original plaster on lath, the floor was hardwood over joists, the ceiling was also plaster. One window, single-glazed, facing the street. One door, solid wood, opening onto an interior corridor. The room had been part of a larger flat that had been converted to offices in the 1960s and was currently unused, which was how she had obtained access.

The acoustic properties of the room were, on paper, unremarkable. The dimensions gave a set of room modes that were well-distributed across the low-frequency range, with no obvious clustering that would produce unusual resonance behavior. The plaster walls would be moderately reflective at mid and high frequencies, absorptive at low frequencies. The single-glazed window was a weak point in the sound insulation, admitting traffic noise and weather, but that was true of most rooms in central London.

And yet.

Mara opened her modelling software and began building a detailed acoustic model of the room. She input the dimensions, the surface materials, the window and door positions. She assigned absorption and reflection coefficients based on published data for plaster, hardwood, and glass. She placed a virtual source at the approximate location where the anomaly signal had been strongest—near the north wall, about a meter from the window—and ran a simulation.

The model predicted what any acoustic model would predict for a room of that size and construction: a reverberation time of approximately 0.8 seconds, a reasonably diffuse sound field at mid frequencies, and a set of room modes that would colour the low-frequency response but would not, in any standard analysis, produce the kind of structured mid-frequency signal she had recorded.

She tried again, adjusting the parameters. She increased the wall reflectivity. She added a hypothetical air leak through the window frame that might produce aerodynamic noise. She modelled the effect of the heating pipes that ran through the wall cavity, checking whether thermal cycling could produce mechanical vibrations at the right frequencies. She spent an hour exploring every physical mechanism she could think of that might produce a signal in the 200 to 3,400 hertz range with formant-like spectral peaks.

Nothing worked. Or rather, everything worked in the sense that she could, with enough parameter manipulation, produce a model that generated energy in the right frequency range. But nothing produced the specific spectral shape she had recorded—the particular pattern of peaks and valleys, the specific formant frequencies, the pulsing temporal envelope that gave the signal its speech-like quality.

This did not surprise her. It did not comfort her either.

The absence of a conventional explanation was not evidence of an unconventional one. It was simply an absence, a gap in understanding that might be filled by better data, better models, better knowledge of the building's construction and history. The history of science was littered with anomalies that had seemed inexplicable until someone found the right explanation, which was always, without exception, a natural explanation, because nature was all there was, and anything that happened in the natural world was by definition natural, however strange it might appear to observers whose understanding was incomplete.

Mara believed this. She had always believed this. It was the foundation of her scientific practice and, in a less articulated way, the foundation of her worldview: that the universe was comprehensible, that phenomena had causes, that those causes operated according to consistent principles that could be discovered through careful observation and rigorous analysis.

But she also knew—had known since the first recording, since the first forty-seven seconds, since the first time she had heard the shaped silence resolve into something that sounded like speech and felt like a hand reaching through a wall—that this particular anomaly was going to test that belief in ways she was not prepared for.

The lab door opened.

"You're here early."

Dr. James Harlan, her department head, stood in the doorway with a takeaway coffee in each hand and an expression of mild surprise. He was a tall man, loosely assembled, with the kind of face that looked perpetually on the verge of asking a question. His research area was underwater acoustics, which meant that his intuitions were calibrated for a medium approximately eight hundred times denser than air, which in turn meant that he and Mara could discuss each other's work with the productive friction of colleagues whose expertise overlapped enough to be useful and diverged enough to be interesting.

"Couldn't sleep," Mara said, which was true.

"Brought you coffee," he said, which was kind, and placed one of the cups on her desk. "I was going to email you, but since you're here. The quarterly review is next Thursday. Have you got something to present?"

"The Fitzrovia surveys."

Harlan nodded. He knew about the Fitzrovia work in its official framing—a systematic study of acoustic anomalies in buildings of a certain age and construction type. He had approved the research allocation and signed the access permits. He had not asked probing questions, because Mara's track record was strong enough to buy her a degree of autonomy that newer researchers did not enjoy.

"Good data?" he asked.

"Interesting data," Mara said. "I've got some anomalous recordings that don't match any standard environmental source. I'm planning a more comprehensive deployment—full array, environmental sensors, multiple sessions over several weeks."

"Cost?"

"Minimal. I've got the equipment. I just need access time and a research assistant for setup and monitoring."

Harlan sipped his coffee. "Who were you thinking?"

"Tomás Vega. He's finishing his MSc project this month and looking for lab work over the summer."

"Good choice. He's careful." Harlan paused. "Anything I should know about the anomalies? Anything that might attract... unwanted attention?"

This was Harlan's way of asking whether her research was going to end up in the tabloids. Acoustic anomalies in old buildings had a way of catching the attention of the paranormal community, and the department had a policy—unwritten but firmly enforced—of not providing ammunition for ghost hunters.

"It's environmental," Mara said. "Almost certainly a building resonance phenomenon. Unusual spectral characteristics, but nothing that can't be explained by the construction. I'm documenting it properly so we can publish without caveats."

"Good." He seemed satisfied. "Send me a summary for the review. And Mara—" He paused at the door. "Get some sleep. You look terrible."

"Thank you, James."

"I mean it. Whatever's keeping you up—"

"It's just the data. You know how it is."

He gave her a look that suggested he knew how it was but also knew that 'the data' was not the whole story. Then he left, and Mara was alone again with her spectrograms and her formant values and her carefully constructed lie, which was not a lie exactly but was a long way from the truth, the way that the surface of the ocean is a long way from the bottom even though they are, technically, the same body of water.

She spent the rest of the morning building the deployment plan. Eight channels, sampled at 96 kilohertz, 24-bit resolution. Environmental logging at one-second intervals. Electromagnetic survey across the spectrum from 50 hertz to 6 gigahertz. Video from two angles. She wrote the protocol document with the care of someone constructing a legal defence, every decision justified, every parameter explained, every potential criticism anticipated and addressed.

She emailed Tomás Vega and asked if he was available for a summer research position. He responded within twelve minutes—the speed of a graduate student who has just been offered paid work—and they arranged to meet the following day to discuss the project.

She did not tell him about the formant comparison. She did not tell him about Lily. She described the project as a methodological study of low-level acoustic phenomena in reverberant environments, which it was, and she described her interest in the Fitzrovia room as architectural, which it had been, and she described the anomalous recordings as promising but preliminary, which was accurate in the way that describing a forest fire as 'warm' was accurate.

At noon, she made herself eat a sandwich from the campus cafe. She sat at a table by the window and watched the students and tried to think about something other than formant frequencies and Mahalanobis distances and the voice of her dead sister saying two words that she still could not—would not—let herself fully hear.

The words were there, at the edge of her perception, the way the anomaly was there at the edge of the spectrogram—a signal that required effort to detect, that could be dismissed as noise if you were inclined to dismiss things, that could be amplified and clarified if you had the right tools and the right patience and the right willingness to listen.

Mara had the tools. She had the patience. She was less sure about the willingness.

Because willingness, in this context, meant something specific. It meant being willing to hear what Lily had said—might have said—and to accept the implications. Not the scientific implications, which were extraordinary enough, but the personal implications, which were worse. Because if Lily was speaking, if some part of Lily persisted in that room, trapped or choosing or simply existing in a way that physics could not explain, then Lily was not gone. Not fully. And if Lily was not fully gone, then the two years of grief and gradual acceptance and hard-won equilibrium that Mara had constructed around the fact of her sister's death were built on a foundation that had just developed a crack.

And cracks, in Mara's experience, did not stay small.

She finished the sandwich without tasting it, threw away the wrapper, and went back to the lab.

The afternoon was spent on the blind analysis protocol. This was the part of the investigation design that she was most careful about, because it was the part most likely to be attacked in peer review, and because it was the part where her own bias—the bias she was trying so hard to contain—was most dangerous.

The protocol worked like this: the full sensor array would record continuously during each session. Mara would not be present in the room during recording. Tomás would manage the equipment, start and stop the sessions, and perform initial data quality checks. The recordings would then be divided into segments and anonymised—stripped of timestamps and session identifiers—before being sent to Mara for analysis.

She would analyse each segment blind, not knowing when it was recorded or under what conditions. She would flag any segments containing anomalous signals, characterise those signals, and extract whatever acoustic features the data supported. Only after her analysis was complete would the segments be de-anonymised and matched to the environmental data.

This was not standard practice for architectural acoustics research. It was standard practice for clinical trials and forensic investigations, where the potential for observer bias was high and the consequences of contaminated results were severe. Mara was aware that the rigour of her protocol might itself attract attention—colleagues might wonder why she was applying clinical-trial methodology to a study of building acoustics—but she judged that the risk of suspicion was lower than the risk of producing results that could not withstand scrutiny.

Because the results, if they confirmed what the preliminary data suggested, were going to attract a great deal of scrutiny. From scientists and sceptics and believers and journalists and everyone in between. And when that scrutiny came, Mara wanted her methodology to be so clean, so transparent, so bulletproof that the only thing left to argue about was the data itself.

At four o'clock, she saved her work and began a fresh analysis of the original recordings. Not formants this time—she had done all she could with formants until she had more data. Instead, she focused on the temporal structure of the anomaly. The pulsing quality she had noticed in the spectrogram.

She extracted the amplitude envelope of the anomalous signal and plotted it against time. The result was a waveform that rose and fell with a quasi-periodic rhythm, not regular enough to be mechanical, not random enough to be stochastic. The peaks were spaced at intervals ranging from roughly 0.2 to 0.4 seconds, with a mean of approximately 0.3 seconds.

0.3 seconds was the average duration of a syllable in English speech.

Mara stared at the plot.

She overlaid the amplitude envelope of Lily's voicemail recording, time-stretched to match the duration of the anomaly. The patterns were not identical—they could not be, given the difference in signal quality and the different words being spoken—but the rhythmic structure was similar. The temporal cadence of the anomaly matched the temporal cadence of natural speech.

This proved nothing. Wind through a crack could produce quasi-periodic amplitude modulation. Mechanical vibration from a loose component could create rhythmic pulses. Even electrical interference from a nearby transformer could generate signals with speech-like temporal properties. The human brain was wired to find speech in noise, to detect patterns in randomness, to hear voices in wind and water and the crackle of fire. Pareidolia was the word for it—the tendency to perceive meaningful patterns in ambiguous stimuli. It was the reason people saw faces in clouds and heard backwards messages in rock songs and believed that the static between radio stations contained the voices of the dead.

Mara knew all of this. She had written about it. She had given lectures on the acoustic basis of auditory pareidolia, had demonstrated to roomfuls of students how easily the ear could be fooled, how readily the brain constructed speech from non-speech, how the desire to hear a thing could create the perception of hearing it.

She knew she might be fooling herself.

But the formant data did not care about pareidolia. Formant values were not a product of perception—they were a physical measurement of spectral peaks, computed by algorithm, expressed in hertz. The algorithm did not want to hear Lily's voice. The algorithm did not grieve. The algorithm extracted frequencies from a signal and reported what it found, and what it found was a spectral signature consistent with a specific human vocal tract, and that vocal tract belonged to someone who was dead.

At five o'clock, Mara locked the drives, shut down the workstation, and put on her coat.

She left the lab with the same careful movements she had used to enter it—closing the door quietly, checking the lock, walking down the corridor with measured steps, as if precision of movement could compensate for the chaos of thought. The building was emptying around her, the day shift giving way to the quiet of evening, and she moved through it like a woman walking through a dream that she knew was a dream but could not wake from.

Outside, the day had turned grey without committing to rain. The sky was the colour of old concrete, low and featureless, pressing down on the city with the gentle persistence of something that had all the time in the world. Mara walked towards the Tube, hands in her coat pockets, mind turning over the same questions it had been turning over since the first recording, the questions that had no good answers, the questions that might never have good answers, the questions that were nevertheless going to structure the next several months of her life.

What was producing the signal? Why did it match Lily's voice? Was the match real, or was she seeing what she wanted to see in noisy data? And if it was real—if more data confirmed the match, if better recordings produced cleaner formant values, if the blind analysis eliminated the possibility of bias—then what?

Then what?

The question had a weight to it that she could feel in her body, in the set of her shoulders and the tightness of her jaw and the way her hands curled into fists inside her pockets. It was not a scientific question. It was not a question that could be answered by data or analysis or rigorous methodology. It was a question about what the world was, about what was possible, about the nature of death and consciousness and the relationship between the physical brain and the thing that the physical brain produced—the self, the awareness, the voice that said I am here and meant it.

Mara descended into the Tube station. The escalator carried her down into the earth, into the warm, stale air of the underground, into the press of bodies and the distant rumble of approaching trains. She stood on the platform and waited and did not think about formants or spectrograms or the Mahalanobis distance between the living and the dead.

She thought about Lily.

Lily, who had been an artist, who had worked in oils and acrylics and occasionally in found materials that she assembled into sculptures that their mother called 'interesting' in a tone that meant she did not understand them. Lily, who had been four years younger and six inches shorter and immeasurably braver. Lily, who had died in a car accident on the M1 on a Tuesday afternoon in November, the kind of ordinary death that happened every day to ordinary people and was no less devastating for its ordinariness.

Lily, who had loved the Fitzrovia flat.

This was the thing Mara had not allowed herself to fully confront. The flat where the anomaly occurred—the flat that she had selected for investigation based on its acoustic properties, based on the reports of previous tenants, based on the age of the building and the construction of the walls and the particular dimensions of the room—was a flat that Lily had visited. Not lived in. Visited. The flat had belonged to a friend of Lily's, another artist, who had used it as a studio. Lily had gone there to paint, sometimes, when her own space was too cluttered or too full of distractions.

Mara had not known this when she selected the site. She had learned it afterward, weeks into the investigation, when the friend—a woman named Priya Sharma, who had mentioned the strange sounds in the room to a colleague of Mara's, which was how the room had ended up on Mara's list in the first place—had mentioned that she had shared the space with Lily.

"She loved that room," Priya had said, her voice careful the way people's voices became careful when they talked about the dead to the dead's family. "She said the light was perfect. She used to work there for hours. Sometimes she'd forget to eat."

Mara had not told Priya about the investigation. She had not told her about the recordings. She had said thank you and changed the subject and gone home and sat in her kitchen and stared at the wall for a long time, trying to decide whether the coincidence invalidated the research or whether the research transcended the coincidence, and finding that the question did not have an answer that she could access from the inside.

Because she was on the inside. She was the investigator and the bereaved and the scientist and the sister, and those roles did not fit together, could not fit together, were as incompatible as wave and particle, except that, like wave and particle, they were all descriptions of the same thing viewed from different angles, and the thing itself—the data, the voice, the signal in the noise—did not care how she described it. It simply was.

The train arrived. Mara got on. She found a seat by the window and watched the tunnel walls flicker past, black and close and lit at intervals by the work lights that gave the underground its particular quality of illumination—not darkness exactly, but an absence of natural light so total that the artificial light seemed to come from everywhere and nowhere, filling the space without warming it.

She would go back to Fitzrovia. She would deploy the full array. She would record and analyse and record again, blind and unblind, with rigour and without mercy, and she would follow the data wherever it led, even if it led somewhere that science had no map for, even if it led to a place where the living and the dead were not as separate as she had always believed, even if it led to a room where her sister's voice still hung in the air like the last note of a song that no one had finished singing.

And if the data, in the end, told her that the anomaly was a building resonance, a standing wave, a trick of architecture and air pressure and the brain's desperate pattern-matching, she would accept it. She would write the paper. She would present the findings. She would add another entry to the long catalogue of things that seemed impossible until they were explained, and she would move on.

But the formant values were in her notebook, written in her own hand, and they matched, and the match was within statistical tolerance, and statistics did not grieve and algorithms did not hope and data did not lie, and the train was carrying her north through the dark, and the question was still forming, and the answer was still waiting, and somewhere in a room in Fitzrovia the air was shaping itself into a voice that the living world could barely hear and that Mara Chen, despite every professional instinct she possessed, could not stop listening for.