Right now
Loading live numbers…
What SEPTA publishes, and what it does not
SEPTA publishes its schedule (static GTFS) and live vehicle positions, trip updates and alerts (GTFS-Realtime) as open data. Buses, trolleys, the Norristown line and Regional Rail all show up with a route, a trip and predictions.
The L and B do not. In the public feeds neither line has a route, a trip or an arrival prediction. Their railcars do appear, reporting a GPS position with a car number and nothing else attached. Whatever a map shows for the L and B beyond those bare dots is built by someone, from those dots and from the published timetable.
This site is not run by SEPTA and has no inside information. Everything here comes from SEPTA's public data.
How L trains are built from railcars
- Find the cars. M-4 railcars carry fleet numbers 1001 to 1220. A vehicle in the feed with no route and a number in that range is treated as an L car. That is an inference from the number, not an official assignment.
- Place them on the line. Each GPS fix is matched to the L's route shape: how far along the line it is, and how far off. A car more than about 60 m from the line, or stopped at a terminal, is in a yard and is not counted as a train.
- Work out direction. From the change in position between fixes, or from the car's heading.
- Group cars into trains. Cars within about 150 m of each other, going the same way, are one train. The train takes the lowest car number as its id.
- Match to the timetable. When a train's position and time fit one scheduled trip closely (within three minutes, and clearly better than the next trip), the map notes the trip and how early or late it is. When it is ambiguous it is left unmatched.
Only a minority of cars report at any moment, and a six-car train may have no reporting car at all. See what coverage means.
The tunnel estimate and how accurate it has measured
Cars report about every 30 seconds while they can see the sky. Underground, the feed keeps repeating the last fix with an ever older timestamp. So when a car that was moving toward a tunnel portal stops getting fresh fixes, the map advances it through the tunnel along the published timetable, station by station, and draws it as a hollow dashed marker. It is never advanced past the far portal. When the car reappears at the other end, the real fix replaces the estimate.
The timetable's tunnel run times turn out to be too fast, so the estimate stretches them by a calibration factor. The factor starts from a value for the train's direction and the time of day, taken from recorded transits, and is then updated from the transits the service observes (it used to start from a constant 1.13 after every restart). Right now: loading…
Measured accuracy: 26 hours of one weekday and the next
Sample size first. These numbers come from 247 car transits (about 167 distinct trains, since cars of one train enter together), recorded from 2:24 PM on Wednesday 11 November 2026 to 4:57 PM on Thursday 12 November, about 26 hours, US Eastern. That is one weekday afternoon through the next afternoon: no weekend, no holiday, one winter day type. It is a much better sample than the first 35-minute recording this page used to quote, but it is still a first measurement and not a promise. The one disruption in it (below) is an anecdote.
| Predicted from | Median error | 90th percentile |
|---|---|---|
| Timetable alone | 122 s | 224 s |
| This tracker's estimate, as it ran (starting factor 1.13) | 41 s | 116 s |
- The timetable alone said the train would surface early in 239 of the 247 transits.
- Observed tunnel time was 0.96 to 1.68 times the scheduled time, median 1.17. The high end is the disruption below.
- Eastbound trains ran about 1.13 times the timetable and westbound about 1.22 (medians over 78 and 80 trains, with the disruption left out). The old single starting factor of 1.13 was close for one direction and low for the other.
- By time of day the median stretch ran from about 1.09 (eastbound, evening) to 1.28 (westbound, 5 to 9 AM). The tracker's starting values are these medians pulled toward the direction's overall value, since each is from few trains; a direction and time of day with fewer than 8 trains falls back to that overall value.
- Fresh test: with the starting values fitted on 11 November only and scored on the 145 car transits of 12 November, the median error was 39 s, against 42 s for the constant 1.13 starting factor with the same online learning. Starting values alone, with no online learning: 47 s against 58 s. These are small differences from a small sample.
- The error is measured against the first fresh GPS fix after surfacing, so it includes the receiver finding its position again.
- The tracker still learns on top of its starting value from the transits it sees; during an inferred delay (below) it does not learn from the slow ones.
Live figure, measured since this service last restarted: loading…
An estimate is not a measurement. The tunnel estimate does not model what is happening on the line: single-tracking, holds and short turns are not in it, and a disruption is only noticed after trains have already run late (see the next section). A train that has not come out well after it should is reported as lost rather than estimated further.
Possible delays, inferred from train timing
When trains take much longer through the tunnel than the timetable says they should, the map shows a notice on the panel: "Possible delay on the L, eastbound: 3 trains are running more than 4 minutes late through the tunnel. Inferred from train timing; not an official SEPTA alert." It is a guess from timing alone. It is not a SEPTA alert, it does not say why, and it can be wrong in both directions: a delay can go unnoticed (it needs at least two trains), and a notice can appear for something that SEPTA would not call a delay.
The rule, per direction: a train counts as late when it is, or ended up, more than 4 minutes past the time the timetable (with the calibration factor) says it should surface. Two different trains late within half an hour start a possible delay. It is cleared only after at least 10 minutes and after two trains have run through normally (under 2 minutes late) with none still late. A single late train never starts one. Only trains that report GPS can be seen, so it can only react to those.
While a possible delay is on, the trains underground in that direction are marked "Possible delay" on the map (with a word and a mark, not just a colour), their estimates are given wider uncertainty, and the slow transits are not used to adjust the calibration factor.
The one example so far
On Thursday 12 November 2026, the 11 car transits (about 7 trains) that entered the tunnel between 4:00 and 5:00 PM ET took a median of 18.3 minutes. The timetable says about 12 minutes, and the usual stretch of 1.17 would give about 14. The first eastbound trains passed the 4 minute mark at about 4:21 PM. Replayed on the recorded transits, this rule would have started a possible eastbound delay at 4:27 PM, and cleared it at 4:57 PM when the second normal train came out. SEPTA's own delay alerts for the L first appeared in its public feed at 4:33 PM (westbound, about an unauthorized person in the track area) and 4:41 PM (a medical emergency at Frankford Transit Center); we do not know the cause of the slow trains. Replayed over the other 25 hours of recorded transits it started no possible delay, with the thresholds above. It was one event, so how often it would raise a false alarm, or miss a delay, is not known. The westbound trains were also slow that afternoon but only one of them passed the 4 minute line, so no westbound delay was inferred.
L tunnel timing
GPS does not work underground, so the tunnel is a blind spot: a car is last seen as it goes in at one end and is seen again, if it reports, as it comes out at the other. The time between is the one thing GPS can measure about the tunnel. These figures count every GPS-reporting L car that went dark entering the tunnel since this service last restarted, so they start from zero after each update and cover only the cars that report.
Loading…
| Direction | Entries | Returned with GPS | Returned late | Not returned | Turned back | Underground now | Dark time, normal returns (min, shortest / median / 90th / longest) | Feed repeated the last fix (median) | Feed dropped the car before it surfaced | First fix past the portal mouth (median) | Late or not returned |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Loading… | |||||||||||
Dark time runs from the last fix before one end to the first fix after the other, so it is a little longer than the portal-to-portal run: GPS is lost shortly before the tunnel mouth and found again 100 to 250 m beyond the far one. A car counts as returned if its first fresh fix comes at or near the far portal within a normal time, returned late if it comes later or farther along, not returned if there is none after about 25 minutes (a car that does reappear later is counted as a new sighting), and turned back if it reappears at the end it entered. A first 75-minute sample (17 entries) had median dark times of 13.6 minutes eastbound (9 cars) and 13.4 westbound (5 cars), and one car that came back about six minutes after its own trainmate; treat small counts as anecdotes.
Why some trains do not reappear
SEPTA's feed repeats a car's last GPS fix for about ten minutes after it goes underground, then drops the car from the feed altogether. A trip through the tunnel takes about thirteen minutes, so a car is missing from the feed for roughly the last three minutes before it surfaces. And some onboard units seem not to resume reporting when the train comes out: in the one case we could check, one car of a train reported again at the far portal while its trainmate stayed silent for about six more minutes. So a train that does not come back with GPS is not necessarily lost or late. On the map it shows as "GPS not returned", at the far portal, until the limit, and then it is dropped.
Time between stations: an estimate
Loading…
| Station | Eastbound, from the west portal (estimate) | Westbound, from the east portal (estimate) |
|---|---|---|
| Loading… | ||
| Trip | Minutes (estimate) | Derived from |
|---|---|---|
| Loading… | ||
Spring Garden is above ground, just outside the east portal, so its leg adds the timetable's above-ground time.
What coverage means
Coverage compares trains the map can see with trains the timetable says are running. "7 of 14 scheduled L trains visible (50%)" means seven L trains have at least one railcar reporting GPS, while the timetable has fourteen in progress. The other seven are running (or should be) but are invisible in SEPTA's public data.
It is a rough ratio, not a precise one. The scheduled count is what the timetable says, and a train running late, early or not at all still counts on one side. An independent open-source project reported, from one Saturday hour of its own, that about half the scheduled trains had no working locator; that is their observation, not a measurement made here.
Right now: loading…
The B, and why it cannot be tracked from public data
Nothing public tracks the B in service. B-IV railcars (fleet numbers 501 to 576 or 651 to 699) report GPS only while they are at Fern Rock yard, the one above-ground point on the line. South of it the line is underground. In a 35-minute recording, 33 B cars reported, 17 of them moved, and none got more than 238 m from the yard.
What the yard does show is cars coming and going. The map watches for a car that was moving through the yard toward the line and then goes silent, and calls that a probable departure into service. A car that reappears coming off the line after a long silence is a probable arrival from service. These are inferences from a few GPS fixes, kept in memory only since the service last restarted, and always labelled probable. A car simply parked in the yard produces no event.
Loading…
For everything else about the B, the only source is the published timetable, plus the estimates described next.
B line estimates
The map can show B trains marked estimated, not measured. Each one starts from a probable departure from Fern Rock (above) and is then carried forward along the published B timetable, station by station, through the layover at the far end and back. The timetable's own trip chains (its block_id field) say which return trip follows which. Nothing is observed once the train goes underground.
No position accuracy has been measured. No B train has been followed from departure to return against a known position, so the uncertainty shown with every estimate is a design choice, not a measured error. Unlike an L train, which resurfaces about 13 minutes after entering the tunnel, a B train gets no correction until its cars come back to Fern Rock, about 45 to 81 minutes later. The estimate is withdrawn once it is no better than the timetable, and a train that does not return on time is marked overdue, with no position.
Local or express. The B runs a local (B1), a weekday express (B2) and a spur (B3). The yard shows only that cars left, not which service they are on. On weekday afternoons a local and an express often leave within a few minutes of each other, and then the pattern is shown as uncertain, with the odds for each. While the two could be in different places no position is drawn at all; the train is listed in the panel only. In a 78-minute afternoon recording, 7 of 13 departures were uncertain this way. That is one afternoon.
Spur trains mostly are not seen returning. A spur train that leaves Fern Rock usually runs the spur for hours and does not come back to the yard in useful time, so its estimate can never be corrected and is withdrawn early.
What it cannot see. Trains that enter service anywhere other than Fern Rock, trains with no GPS-reporting car, and anything between departure and return: delays, holds, short turns, swaps. These estimates are not in any GTFS-Realtime output.
Loading…
Scheduled trains, and how they differ from tracked ones
The published timetable is the only source that covers every B train and the L trains that have no reporting car. For any moment, the map works out where each trip in progress should be by interpolating between its scheduled stop times along the route shape, using the right service for the day (weekday, weekend or holiday exceptions, in Eastern time, including trips that run past midnight).
On the map these appear as small, pale train squares with a dashed grey outline and the line letter, under the heading "Scheduled trains (timetable, not tracked)", with separate switches for the B and for L trains that are not tracked. Both are off by default.
| Tracked L train | Scheduled train | |
|---|---|---|
| Marker | Blue "L" train square with a direction notch (solid for a GPS fix, hollow dashed for an underground estimate) | Small pale train square with a dashed grey outline |
| Source | A railcar's GPS, plus the timetable underground | The timetable only |
| Means | A railcar of this train was seen | This is where the timetable puts the train; it may be early, late or not running |
A scheduled position is not an observation. SEPTA publishes no real-time data for these trains.
Songs of SEPTA (preview)
An experimental, optional way to listen to the map. It is a preview: it is not on the page's controls, and it only exists if you add ?songs=1 to the map's address.
- What the sounds mark. A sound marks an animated passage: the drawn marker of a bus, trolley or train passing a stop or a meeting place, arriving at the end of its route, or passing another drawn marker. The markers glide between fixes, so these are passages of a drawn marker, not observations that a vehicle stopped, arrived or met another. Amtrak is the one exception to the gliding: a note sounds when Amtrak's reported position is close to a station it shares with SEPTA, and only while that reported position is under eight minutes old. Amtrak's positions usually reach this page a few minutes old, so each such sound is labelled as a reported position and states its age, for example "reported position, 5 min old".
- Estimates are muffled and labelled. Markers drawn at a measured GPS position sound in full. Estimated B trains, the B's timetable positions (only while that layer is on) and, if you switch them on, estimated L trains underground sound muffled, as if heard through a wall, and are labelled as estimates or timetable positions on screen. L timetable positions never sound.
- The harmony is composed; the timing comes from the city. The chords are written ahead of time and cycle slowly. Every note a vehicle plays is a tone of the chord in force, so any mix of vehicles sounds consonant. What comes from the data is when notes happen, where, and which kind of vehicle makes them, not the tune.
- Some passages are left out. There are far more passages than notes can carry, so a limit on how many sound per second applies. The list of recent events in the panel shows the passages that were shown and whether each one sounded or was muted. Passages the rate limit leaves out never appear in that list; they appear only in the panel's per-class counts (raw, eligible and shown).
- Sound waits for your first gesture. Opening the map with
?songs=1arms Play. Sound starts on your first click or key press (a browser rule), or at once if your browser allows autoplay. Once playing it continues in a background tab until you press Pause, which holds it until you press Play again.
Open data outputs
Everything on the map comes from public endpoints you can use too. They are read-only, cached for a few seconds, and served from memory; a visit to this site never makes a request to SEPTA.
- JSON API. Interactive documentation. Useful endpoints:
/api/v1/vehicles,/api/v1/heavy-rail/l/trains,/api/v1/heavy-rail/scheduled?line=L(orB),/api/v1/heavy-rail/b-yard,/api/v1/coverage. - MCP server for AI tools, at
/mcp(streamable HTTP). It offers the main read tools, including L and B trains, tunnel timing and scheduled trains, but not every API endpoint. - GTFS-Realtime feed of tracked L trains:
/gtfs-rt/l/vehicle-positions.pb. GTFS-Realtime has no way to say "estimated", so treat it as an unofficial estimate; add?format=jsonto read it in a browser.
Credits
- Vehicle positions and the timetable are SEPTA's public data. This is an unofficial use of it.
- Visitor analytics: on the public site the map, this page and the on-time page load Google Analytics to count visits and see which pages are used. It sets one first-party Google Analytics cookie (_ga) so that a returning visitor is counted once; visitors in the EEA, the UK and Switzerland get Google's cookieless mode instead, with no cookies. No advertising cookies are set. The API, the ride logger and the MCP endpoint do not load it.
- Advertising: on songsofsepta.com the map shows listings from Strommen Systems' own surplus electronics shop, which pays for this site, and pages may load Google AdSense. AdSense runs with ad storage denied, so it serves non-personalised ads and sets no advertising cookies.
- Vehicle photos: the popups show third-party reference photos from Wikimedia Commons and Flickr, under Creative Commons licences or in the public domain. They are cropped. The photographer, title, source and licence are shown beside each photo and listed at /api/v1/fleet-photos. A photo marked as an example is not the vehicle on the map. Photos are archive pictures, not live images.
- Map tiles and data are © OpenStreetMap contributors.
- Prior art: garlicbuffalo/septa-transit-alerts showed that L trains can be built from the unrouted railcars, and that scheduled tunnel run times need stretching by a learned factor. The grouping of cars into trains and the original starting value of 1.13 follow that project's approach. The code here is independent.