The Last Coda

Keza Paredes had not played the clip in four months. She had put it away in March, after the permit was denied and the grant committee had asked for a preliminary finding and she had given them one that was technically true and substantively empty. She took it out now because the matriarch was three hundred meters off the bow and the hydrophone was in and the clan was clicking and Keza had run out of ways to pretend she was not listening for the same nine seconds she had been listening to for seven years. The clip was nine seconds. The first four were her father’s codas, recorded in 1998 off the coast of Dominica, and the last five were the matriarch’s reply. The reply began with his first coda’s rhythm, then bent it, then bent it again, and ended on a pattern Keza had not heard in any other recording in fifteen years of listening.

The matriarch was Marengo, SW-07 in the Project CETI catalog. A female sperm whale roughly forty years old, leader of a clan of eleven that worked the waters east of Dominica. Keza had first recorded her in 2019, off the coast where her father had died. She had followed Marengo since. Seven years of hydrophone drops, spectrogram reads, and the slow accumulation of a picture that was either a breakthrough or a very expensive mistake.

Keza’s father had been a fisherman. He had drowned in 2019 when his boat capsized in a squall off Scotts Head. He had been alone. The coast guard found the boat two days later. They found him the next morning, tangled in his own net, three hundred meters from shore. He had been fifty-one.

Keza had been twenty-six. She had been in Boston, finishing her master’s at MIT, when the call came. She had flown home and stood on the dock at Roseau and listened to the water and heard nothing. She had gone back to Boston and finished the degree and then come back and spent seven years learning to hear.

Her father had been a fisherman who talked to whales. Not in any scientific sense, though he had recorded them: four codas on a cassette in 1998, with a hydrophone hung over the side of his boat. He had sat in his boat and clicked his tongue at them, and he had believed, with the unshakeable certainty of a man who had grown up on the water, that they clicked back. Keza had grown up thinking this was a story. A fisherman’s story. The kind of thing a father tells a daughter on a boat at sunset.

Then she had recorded Marengo, and Marengo had clicked back, and the story had stopped being a story.

The vessel was the Rosalind, the fourteen-meter research catamaran Keza had run for seven years. Keza was the senior bioacoustician. Her funding ended in November. The department at UWI wanted a result. She had a second paper, built on the first, in review at Nature, claiming the codas were translatable. The paper had been in review for eleven months. The reviewers wanted one clean exchange. One turn-taking event that proved the model worked. Not a statistical pattern across hundreds of recordings. One exchange. One moment where the whale said something, Keza said something back, and the whale replied in a way that could not be explained by chance or by the model’s own assumptions.

Keza had two hundred and seventeen transmissions in the archive. She could have run the analysis on her laptop in an afternoon. She told herself she needed fresh data. She told herself the archive was contaminated by her own expectations. Both of these things were true, and neither was the real reason.

The trial was simple. Each morning at 06:14, she would play a coda through an underwater speaker. The coda was five clicks, pattern 1+1+3, recorded from Marengo in 2019. Then she would listen. If Marengo replied within a window that matched the model’s predicted turn-taking latency, the exchange was validated. If not, it wasn’t.

The model predicted turn-taking at a mean latency of 2.3 seconds, with a standard deviation of 0.4. Marengo’s clan had a turn-taking signature that was distinct from other clans. Keza had published this in 2023. The paper had been cited forty-one times. It was the foundation of her career.

The problem was that the model predicted turn-taking. It did not predict what the whale would say. It predicted when, not what. Keza had been claiming, in the paper under review, that the codas were translatable. That the content of the turn could be decoded. That the five clicks meant something specific, and that the reply meant something specific, and that the exchange was a conversation.

She had not proved this. She had proved that whales took turns. She had dressed up turn-taking as translation and sent it to Nature.

The editor had asked for one exchange. One moment where the turn-taking was paired with content. One reply that was not just a reply, but an answer.

Keza needed that exchange. The money ran out in November. The department wanted a result. The paper was her career. And underneath all of that, underneath the funding and the tenure track and the department chair’s emails, there was the other thing. The thing she did not say out loud. She wanted Marengo to talk to her. Not to the speaker. Not to the hydrophone. To her.

She had predicted the clan’s behavior forty-eight times in a row, and knew their dive patterns, their foraging depths, their social grouping. She knew that Marengo’s clan used a coda type that the Dominica Project called 5+1, five clicks followed by a longer pause and a sixth. She knew that other clans in the Caribbean used different coda types. She knew that the coda type was clan-specific, a social marker, like an accent.

She did not know what the clicks meant. She had built a model that predicted when whales would click, and she had called it understanding, and she had sent it to Nature.

The Rosalind sat at anchor in the lee of Dominica, in water three hundred meters deep. The hydrophone was in. Keza could hear the clan clicking through her headphones. The clicks came in bursts, irregular, like a conversation she could hear through a wall. She could not make out the words. She could not even tell if there were words.

She closed the laptop. She put on her shoes. She went up on deck.

The morning was gray. The Atlantic was flat. Dominica was a green wall to the west, the mountains lost in cloud. The Rosalind rocked gently at anchor. Keza could see Marengo’s blow two hundred meters off the bow, a slow white column against the gray water. The clan was close. They had been close all week. Keza did not know if this was luck or if Marengo had learned to associate the Rosalind with something. She did not know what.

She went below and checked the hydrophone. The signal was clean. She checked the underwater speaker. It was positioned at fifteen meters depth, aimed at the clan’s usual approach bearing. She loaded the playback file. Nine clicks, recorded from Marengo in 2019. The same coda she had been playing all week.

She checked the time. 06:12. Two minutes.

She sat on the cabin floor with the laptop on her knees. The cabin smelled of diesel and old coffee. She put on her father’s headphones. The headphones were old, the foam cracked, the cable patched with electrical tape. Her father had used them for fishing. He had used them to listen to the radio. He had never used them to listen to whales. Keza had taken them after he died because they were his, and because they were the only thing of his she wanted.

06:13. She watched the clock.

At 06:14:00, she pressed play.

The coda went out through the speaker. Nine clicks, sharp and clean, cutting through the water. Keza heard them on the hydrophone a fraction of a second later, the speaker’s output picked up by the receiver. Then silence. Then the clan’s ambient clicking, the background noise of sperm whales going about their business.

She watched the spectrogram. The screen showed frequency on the vertical axis and time on the horizontal. The coda appeared as a series of sharp vertical lines, each one a click. They were clean. The playback was clean. She had checked the calibration three times.

She waited. The model predicted a reply at 2.3 seconds, plus or minus 0.4. She watched the clock. One second. Two seconds.

At 2.1 seconds, a coda came in on the hydrophone.

Keza looked at the spectrogram. Nine clicks. The same coda type. The same rhythm. But the last click was forty milliseconds longer than the playback’s last click. The same elongation she had measured in the 2019 recording. The same pattern.

She saved the file to three locations. She sat on the cabin floor and looked at the spectrogram and felt nothing she could name.

She played it again. On the sixth listen, she noticed something she had not seen before. She had been looking at the whole coda. The rhythm, the timing, the elongation. She had not looked at the individual clicks.

She zoomed in on the spectrogram. The first click of Marengo’s reply was not a single click. It was a doublet. Two clicks, separated by twelve milliseconds. Keza had not seen this because she had been looking at the coda as a whole, as a unit. The doublet was inside the first click, buried in the onset, invisible at the resolution she had been using.

She changed the resolution. The doublet appeared clearly. Two clicks, twelve milliseconds apart, at the start of the first click. She had seen this pattern before. She had seen it in the 2019 recording. She had seen it in other recordings from Marengo’s clan. She had catalogued it. She had a name for it. She called it a prefix marker. She had assumed it was a clan identifier.

She had assumed this because the doublet appeared at the start of codas produced by Marengo’s clan and not at the start of codas produced by other clans. It was a clan signature. It was identity.

But it was not at the start of every coda. It appeared only in certain codas. Codas that were produced in certain contexts. Keza had noted this. She had not pursued it. She had filed it under clan-specific variation and moved on.

She looked at the doublet in Marengo’s reply. It was there. It was there in the playback’s coda too, because the playback was recorded from Marengo. But in the reply, the doublet was different. The interval was not twelve milliseconds. It was fourteen.

Keza stared at the screen. Fourteen milliseconds. Two milliseconds longer than the clan’s standard doublet. Two milliseconds that her model did not predict, did not account for, did not have a category for.

She checked the archive. She pulled up every recording of Marengo’s reply codas over seven years. She looked at the doublets. In every case where Marengo had replied to a transmission from the Rosalind, the doublet interval was fourteen milliseconds. In every case where Marengo had produced a coda without a preceding transmission, the doublet interval was twelve.

Twelve milliseconds when speaking to the clan. Fourteen when speaking after the Rosalind spoke.

Keza sat on the cabin floor and looked at the numbers on her screen. Fourteen. Twelve. The difference was two milliseconds. It was smaller than the resolution of most bioacoustic studies. It was smaller than the margin of error in her own measurements. But it was consistent. Seven years of consistency. Every time the Rosalind spoke, Marengo’s reply doublet shifted by two milliseconds.

Keza had a model. The model predicted turn-taking. It did not predict doublet intervals. It did not predict anything about the content of the clicks. It predicted when, not what, and it predicted when well. It did not predict fourteen milliseconds.

She ran the numbers again. She checked the calibration. She checked the hydrophone’s clock drift. She checked the speaker’s latency. Everything was within spec. The fourteen milliseconds was real.

She sat with it for a long time. The cabin was quiet except for the hum of the generator and the faint sound of water against the hull. Through the headphones, she could hear the clan clicking. The clicks came in bursts, irregular, like talk heard through a wall.

Then the ambient clicking changed. The bursts were longer. The pauses between them were shorter. The rhythm was different.

She pulled up the spectrogram of the ambient channel. The clan was producing codas. Not the 5+1 type. A different type. One she had catalogued but rarely seen. It was a type that appeared in the archive only in certain contexts. Contexts she had not been able to identify.

She listened. The codas were coming in a sequence. One after another, with pauses between them that were not random. The pauses were structured. They were turn-taking pauses. But the turns were not between Marengo and the Rosalind. They were between Marengo and other members of the clan.

Marengo was having a conversation with her clan. And the conversation had started when the Rosalind spoke.

Keza played the recording back. The Rosalind’s playback coda went out. Two seconds. Marengo’s reply came in, with the fourteen-millisecond doublet. Then, three seconds later, another coda from Marengo. Then a reply from another whale. Then Marengo again. Then another whale. The sequence continued for four minutes.

She had never heard this before. In seven years of recording, she had never heard Marengo initiate a clan conversation immediately after replying to the Rosalind. The conversations happened. She had recorded them. But they had always seemed spontaneous, tied to foraging or social events. This one was tied to the Rosalind’s transmission.

Keza pulled up the archival recordings. She searched for instances where a Rosalind transmission had been followed by a clan conversation. She found nine. In all nine, the conversation had started within thirty seconds of the reply. In all nine, the conversation had used the same coda sequence. In all nine, Marengo’s first reply had contained the fourteen-millisecond doublet.

She had not seen this because she had been looking at the reply. She had been looking at the exchange. The turn between the Rosalind and Marengo. She had not been looking at what happened after.

What happened after was that Marengo talked to her clan. And her clan speech differed from her address to the Rosalind. The doublet was twelve milliseconds in clan speech. Fourteen in speech directed at the Rosalind. The coda types were different. The rhythm was different. The pauses were different.

Marengo was not replying to the Rosalind. She was reporting to the Rosalind. She was telling the clan something about the Rosalind. Or she was telling the Rosalind something about the clan. Or she was doing something that Keza’s categories could not hold.

Keza sat on the cabin floor and looked at the spectrograms. She had spent seven years building a model that predicted turn-taking. The model worked. It predicted when whales would click. It did not predict what they were doing when they clicked. It did not predict the fourteen-millisecond doublet. It did not predict the clan conversation that followed. It did not predict any of this because the model was not a model of communication. It was a model of timing.

She thought about her father. He had clicked his tongue at whales and believed they clicked back. He had been right. They had clicked back. But he had not known what the clicks meant. He had not needed to know. He had clicked, and they had clicked.

Keza was not fine with that. Keza had spent seven years trying to get to the meaning. She had built a model. She had published papers. She had sent a paper to Nature claiming the codas were translatable. And the codas were not translatable. They were not translatable because they were not words. They were not words for things. They were acts in a social world that Keza could observe but not inhabit.

The doublet was not a prefix marker. It was an address. Marengo was addressing the Rosalind. Not the speaker. Not the hydrophone. The Rosalind as a social actor. She had been addressing the Rosalind for seven years. She had been addressing Keza. And Keza had been measuring the address as data.

Keza pulled up the 2019 recording. The one that started everything. The one with her father’s codas and Marengo’s reply. She looked at the doublet in Marengo’s reply. Fourteen milliseconds. Marengo had been addressing the Rosalind since the first day. Since the first transmission. Since Keza’s father died and Keza came back and started listening.

The reply began with his first coda’s rhythm, then bent it, then bent it again, and ended on the elongated click. Keza had thought the reply was an answer. It was not an answer. It was an address. Marengo was not saying “I heard you.” She was saying “I am speaking to you.” That distinction was everything.

Keza closed the laptop. She sat on the cabin floor. The headphones were around her neck. Through the hull, she could hear the clan clicking. Not through the hydrophone. Through the hull. The sound traveled through the water and through the steel and through the air and into her ears, and it was sound, not data, and she heard it.

She went up on deck. The morning was still gray. Marengo’s blow was still visible off the bow. The clan was still there. Keza stood on the deck and listened. The clicks came through the water and through the hull and through the air, and they were not codas. They were not data. They were sound. They were the sound of animals talking to each other about something that had nothing to do with her.

She went back below. She sat at the laptop. She opened her log file. She typed:

“They are not talking to us. They are talking near us.”

She saved the file. She closed the laptop. She put on her shoes. She went up on deck.

The hydrophone was still in the water. The speaker was still in the water. The recording equipment was running. The data was flowing. Keza could hear it through the headphones, the clicks coming in as sharp vertical lines on the spectrogram, as numbers in a file, as evidence for a paper that claimed to understand.

She reached down and turned off the hydrophone.

The whales clicked. The water moved. The boat rocked. Keza stood on the deck and listened, and the sound came through the hull and into her ears. She did not turn the hydrophone back on.