ADR-0048: The Turntable Is Heard on Both Channels, and One Gone Silent Is Said on the Console

Context

On the evening of 2026-10-25 (about 01:34 UTC on the 26th) the stand went to standby in the middle of a side. A record was playing, the Eagles' Live, and the music could still be heard in the room. The turntable’s driver had heard twenty seconds below -55 dBFS and said stopped, as ADR-0026 has it do, and nowplayd dimmed the room a second later.

Nothing else had failed. The driver logged no turntable.capture_lost: the sound card was there and sending audio. The recognizer was well: the minute’s counts for the Shazam endpoint said limited=0 failed=0 (ADR-0027).

The input had lost its left channel. The driver hears the amplifier’s tape output, Tape 1 out on an Onkyo A-8170 integrated amplifier from the mid-1980s, whose aged tape selector contacts can drop a channel. A stereo meter on the same card, arecord -c 2 -V stereo, read +00%|76%: the left side silent and the right side carrying the record. After the Tape 1 selector was worked back and forth and set back to phono it read 29%|26%.

The driver listens to one channel, and it was the silent one. It asks arecord for one channel (-c 1) from plughw:CODEC, so that ALSA’s plug layer converts the card’s stereo to mono. What was seen is consistent with that conversion taking the first channel, the left, and dropping the right, rather than mixing the two: with the right channel at 76% the driver heard silence. That is behavior observed on this card and this ALSA, and is written down as such; no ALSA document has been checked for it.

The console showed nothing wrong. Its “Song detector” pill was green throughout. That is what ADR-0034 says it means: the sound card is heard, not that the turntable is on, and not that both channels arrive. A person looking at the console saw a green pill, a dim stand and music playing, and nothing to say which part was wrong.

Options considered

What the driver listens to

One channel, as now. Nothing changes, and a fault in one channel of a cable that is known to be intermittent stops the stand whenever it is the first channel, and is unseen whenever it is the second.

Both channels, mixed to one, and nothing said. -c 2, and the two added together before anything is measured. The stand keeps working with a channel gone, about 6 dB quieter, which is well above -55 dBFS for music. The fault is hidden: the record sounds one-sided in the room and nothing in EdS says so, and the next time it is noticed is when the other side goes too.

Both channels, each measured, and the louder one decides. The stand keeps working with a channel gone, and the driver knows which one and can say so. It costs a stereo capture, 64 KB a second from the card instead of 32, and a second level per tenth of a second.

Where a silent channel is said

A richer status. online, offline, and a third word such as degraded. One topic, already on the console. But a one-sided input is not offline, and the status topic is not only for people: nowplayd’s arbiter reads a source’s status, and the broker sends its last will there. A third word would have to be taught to every reader of presence, and a reader that does not know it would take it for something else.

In the source’s state. A field on nowplaying/source/turntable/state. The state is the contract every source shares (ADR-0006), and a channel is not something any other player has. A change of channel would also be a change of state for nowplayd to compare and pass on.

A retained topic of its own, beside the status. One more topic and one more grant on each side of it. It says one thing, and only those who ask for it hear it.

An alert beyond the console: a notification, or a dashboard panel with a threshold. Not now: the console and the event in the log store are where this would be looked for, and a phone being told about a cable is more than the fault deserves.

Decision

The turntable’s driver captures both channels and measures each. Music is decided by the louder channel, so one dead channel no longer stops the stand. A channel that stays silent while the other carries music is published on a retained topic of its own, logged once, and shown on the console.

• Both channels are captured (arecord -c 2, interleaved, through the same plughw device), and each tenth of a second is measured once per channel with the same Level.

• A block is music when the louder channel is at or above the silence level (NP_TURNTABLE_SILENCE_DB, -55 dBFS). Everything ADR-0026 decides from silence, the gap and the stop, is decided from that. Its timings do not change.

• The clip sent to the recognizer is the two channels mixed, or, while one channel is reported silent, the other alone, so that a dead channel does not halve what the recognizer is given.

• A channel is silent when it has been below the silence level for the stop time (NP_TURNTABLE_STOP_AFTER, 20 s by default) while the other has carried music. It has returned when both have carried music for the same time. The same twenty seconds each way is what keeps a contact that crackles from saying so twice a second. While neither channel carries music, between sides or with the record lifted, nothing can be told about either, and the word stands as it was.

• It is published, retained, on <prefix>/nowplaying/source/turntable/input, beside the source’s status, when the word changes and at no other time:

  {"v":1,"silent":"left","left_db":-90.3,"right_db":-24.1,"changed_at":"2026-10-26T01:34:12Z"}

  silent is left, right or none; the levels are each channel’s over the last second when the word changed, for whoever reads it in mosquitto_sub; changed_at is the driver’s clock, as on a state. Field names follow the contract’s (snake_case, v first). The topic is published none each time a capture starts, so a stored word from before a restart does not outlive it.

• The topic is the turntable’s only. Other drivers have no input to measure, and publish nothing there.

• It is logged once each way, as turntable.recognizer_failed is once for a run of failures: a warning with the event turntable.channel_silent and which channel when the word becomes left or right, and an info line with turntable.channel_back when it returns to none. Each is published to the agent’s log topic like the agent’s other events (ADR-0017).

• The console’s “Song detector” pill goes amber, with “left channel silent” or “right channel silent” beside it, while the topic says so and the driver’s presence is green. A driver that is gray or red is shown as ADR-0034 has it, whatever its input last said: a stored word about a channel is no more believed than a stored online.

• nowplayd does not read it. Its subscription, nowplaying/source/+/+, already delivers the topic, and the hub ignores a source topic that is not state, status or art; the arbiter decides nothing from a channel.

• The console still only looks (ADR-0018). It is granted one more subscription and nothing more to publish.

Consequences

• A record keeps the stand lit with one channel gone, and is named as before: the recognizer hears the live channel.
• A fault the room can hear is now one a person can see, on the console within twenty seconds of the music, and in the log store as one line for each time it happens.
• Two grants, in the tenant’s Terraform and in the backend’s copy of the accounts (images/backend/files/accounts.conf): np-relay publishes nowplaying/source/turntable/input, and console subscribes to nowplaying/source/+/input, added to view.Topics as written. The console’s account is not granted nowplaying/source/+/+; it is granted each kind of source topic by name, and the broker refuses a subscription it was not granted without saying so. Applying them is a change of its own, before the driver and the console that use them are deployed.
• The capture interface changes from mono to stereo. The driver’s tests that feed it audio feed two channels, and gain cases for each channel going silent and coming back, and for both going silent at the end of a side.
• A record with a channel truly silent for twenty seconds (a passage hard to one side, or a mono record played through a cartridge wired to one channel) is shown amber until both channels carry music again. That is said, not acted on: the stand is not dimmed for it.
• A stereo capture is twice the bytes from the card, 64 KB a second instead of 32. On the agent Pi that is nothing.
• A dead channel is still a dead channel. EdS says which one; the contact is still cleaned by a person.
• What the plug layer does with -c 1 stops mattering here: the driver no longer asks it to choose.

What would settle it

Not yet shown:

• The driver’s tests, with stereo audio: a side with the left channel silent throughout is followed to the end and named; the word becomes left after twenty seconds and none after twenty seconds of both; a side’s end with one channel silent leaves the word as it was.
• On the agent Pi, arecord -c 2 through plughw:CODEC delivers both channels interleaved, left first, as the stereo meter shows.
• A side played with one side of the tape output unplugged: the stand stays lit, the track is named, input says which channel within about twenty seconds, and one turntable.channel_silent reaches the log store.
• The plug put back: input says none, one turntable.channel_back is logged.
• The console’s “Song detector” pill amber with the channel named while that is so, green after, and gray, not amber, while the driver is not heard.