ElevenLabs text-to-speech in Galleria: end-to-end test evidence

What each hop was shown to do

• exercised

  Galleria UI

  Controls, live counter, live quote, over-limit block, playback, captions panel
• exercised

  Gateway

  400 before the credit gate over the limit; fal dispatch; quote from the prompt length
• exercised

  go-queue

  One start and one completed line per generation; failures archived, not retried
• exercised

  vertex-ai to fal

  Only the model's own keys forwarded; Auto language sends nothing
• exercised

  Storage

  MP3 stored and served; word timings saved on the asset
• exercised

  go-credits

  Cost-plus per character; one usage row per generation
• exercised

  Public API

  Neither new model is listed; Seed Audio and every other model as before
• exercised

  MCP list_models

  Called over HTTP: same result as the public list

Quote equals charge

For every completed generation the credits charged in credit_ledger equal what the estimate returned for the same character count, on the cost_plus pricing path, with exactly one usage row.

Four more short generations were each quoted and charged 1 credit: a 22-character Japanese sentence and a 47-character 192 kbps run on v4, and one 39-character run per model and a final 38-character v4 run after the scope passes. Turbo accepted 5,000 characters, which fal does not document; bytes per character matched the 1,013-character run, so nothing was truncated.

In the browser

The two generations made in the browser

Prompt for both: "The tide is turning now. Bring the small boats in before dark." (62 characters). Press play to hear what was stored.

Request bodies captured from the browser

POST /api/generations  ->  201   (Eleven v4)
{"user_id":"7c1e96fa-…","model_id":"4b89989b-…","prompt":"The tide is turning now. Bring the small boats in before dark.",
 "config":{"voice":"Aria","stability":0.5,"similarity_boost":0.75,"language_code":"","apply_text_normalization":"auto",
           "output_format":"mp3_44100_128","timestamps":true},
 "asset_type":"audio","action":"generate","organization_id":"036bb424-…"}

POST /api/generations  ->  201   (Eleven Turbo v2.5)
{"user_id":"7c1e96fa-…","model_id":"2263156a-…","prompt":"The tide is turning now. Bring the small boats in before dark.",
 "config":{"voice":"Rachel","stability":0.5,"similarity_boost":0.75,"style":0,"speed":1,"language_code":"",
           "apply_text_normalization":"auto","timestamps":true},
 "asset_type":"audio","action":"generate","organization_id":"036bb424-…"}

Live quote requests

POST /api/credits/estimate   {"model_identifier":"eleven-v4-tts", … ,"characters":62,   "config":{…}}
  -> {"credits":"1","cogs_usd":"0.00496","pricing_path":"cost_plus","estimated":false}
POST /api/credits/estimate   {"model_identifier":"eleven-v4-tts", … ,"characters":1000, "config":{…}}
  -> {"credits":"6","cogs_usd":"0.08","pricing_path":"cost_plus","estimated":false}
POST /api/credits/estimate   {"model_identifier":"eleven-turbo-v2.5-tts", … ,"characters":1000, "config":{…}}
  -> {"credits":"4","cogs_usd":"0.05","pricing_path":"cost_plus","estimated":false}
Empty prompt (no characters key)
  -> {"credits":"1","pricing_path":"flat_fallback","estimated":false}

What vertex-ai sent to fal

Auto language put an empty language_code in the request, and vertex-ai dropped it: neither line below carries the key.

[FAL Audio] model=eleven-v4-tts path=elevenlabs/tts/eleven-v4 chars=62
  params=map[apply_text_normalization:auto output_format:mp3_44100_128 similarity_boost:0.75 stability:0.5 timestamps:true voice:Aria]
[FAL Audio] stored audio path=gs://goodtake_ai_dev/audios/fal-tts/fal_tts_91851304_20261004_162728.mp3 content_type=audio/mpeg

[FAL Audio] model=eleven-turbo-v2.5-tts path=fal-ai/elevenlabs/tts/turbo-v2.5 chars=62
  params=map[apply_text_normalization:auto similarity_boost:0.75 speed:1 stability:0.5 style:0 timestamps:true voice:Rachel]
[FAL Audio] stored audio path=gs://goodtake_ai_dev/audios/fal-tts/fal_tts_c8fd8df8_20261004_163025.mp3 content_type=audio/mpeg

Database

Concatenating the 12 stored word values of either asset gives the prompt back exactly. The go-credits spend lines for both read path=cost_plus; go-queue logged one start and one completed line per id.

Downloaded caption files

1
00:00:00,000 --> 00:00:01,720
The tide is turning now.

2
00:00:01,920 --> 00:00:04,080
Bring the small boats in before dark.

WEBVTT

00:00:00.000 --> 00:00:01.720
The tide is turning now.

00:00:01.920 --> 00:00:04.080
Bring the small boats in before dark.

1
00:00:00,070 --> 00:00:01,277
The tide is turning now.

2
00:00:01,486 --> 00:00:03,181
Bring the small boats in before dark.

1
00:00:00,000 --> 00:00:02,320
今日は、良い天気ですね。

2
00:00:02,320 --> 00:00:03,520
散歩に行きましょう。

Driven through the API

Scope passes

The owner's rule: leave existing behaviour alone unless the new models needed it changed, keep the implementation simple, and remove workarounds. Three independent audits of the diff, and a line-by-line check by the reviewer, led to these changes. All are re-verified above.

• Back to exactly what main has: the audio success toast, two comments in the pricing engine, the public API doc page, the Audio feed's empty-state sentence, a startup log line, Seed Audio's public listing, and all of Seed Audio's prompt-limit code (its check, its message, its counter and its config entry).
• The new models carry their own limit. Their counter and over-limit check sit beside Seed Audio's rather than replacing it. Two existing conditions gained one term each, so the old "3000" blocks do not show on a 5,000-character model.
• The public list hides only the two new models. An earlier version rejected the whole audio type, which also delisted Seed Audio.
• Removed from the new code: a second null filter in the queue worker (the gateway already filters once), and an error type in vertex-ai whose 400-or-500 distinction nothing read.

Minimal-diff audit, 6 October. A fresh independent reviewer read every changed hunk against main. Two things touched existing material without need and were undone: a helper that had been extracted out of Seed Audio's enqueue function (that function is byte-identical to main again; the new function carries its own copy of the call, as the image and video ones do), and another ticket's route (/dashscope/generate-video) that the regenerated API docs had picked up. The panel now counts characters only for models that carry their own limit, so image, video and Seed Audio renders run none of the new counting code. After the changes one live Turbo generation was charged exactly its quote, and the panel counted and quoted as before in both layouts.

Audit of the added code, 6 October. A second independent reviewer classified every added line as required, convention, optional or redundant. Removed as unneeded: a frontend flag that always travelled with the text limit, a duplicate unknown-model check in the speech route (the service already rejects it), and most of two documentation additions. Kept on purpose: the admin price preview, the quoted amount on the ledger, the splitting of Japanese and Chinese captions, the tests that are the only coverage of a contract, and the Language, Text Normalization and Output Format settings, which the ticket lists. The small test that pins which models are Galleria-only was removed in this pass and restored the same day, because the ticket asks for it. After the removals the panel counted and quoted exactly as before, and one live generation with timestamps was charged its quote.

The change set now modifies 17 lines that existed before it; everything else is addition. Each of the 17 is forced: four trailing commas, the audio handler's provider reject and enqueue call and two comment lines, the credit-gate call, the admin unit list with its message and one comment line, the pricing doc's unit list, one import, two counter conditions, and the Projects chat's model filter.

One difference remains for existing models, and no user can see it: the gateway sends the prompt's character count on every pricing request, not only for the new models. The engine reads it only for per-character rows. An image model's estimate was byte-identical with and without the field on all three pricing paths.

Is it Galleria only?

Yes, by the owner's definition: outside Galleria the two models cannot be selected in any picker and are not listed by the Claude connector. The Projects chat picker was the one place that listed them; it now hides these two models and nothing else. The last three rows are not places where a model is picked or listed for a user, and are left as on main.

Two results that were not a clean pass

• Eleven v4's last caption cue ends 31 ms after the audio. The cue ends at 4.080 s; ffprobe measures the MP3 at 4.049 s. The files reproduce fal's word timing faithfully, and the same 31 to 46 ms overrun shows on every v4 asset measured, while Turbo assets stay inside. This is a difference between the provider's alignment data and the encoded file length. Players end the cue with the media.
• The picker's option list has no per-model logos. This is how the shared picker already renders for image and video models. The ElevenLabs mark loads correctly on the selected model's card.

Not tested

• Audible output from speakers. Playback was established from the audio element's state, and the stored MP3s are above.
• A browser generation with Timestamps off, 192 kbps, a named language, or non-default sliders. Those values were seen in quote requests and exercised through the API, not generated from the UI.
• The over-limit block on Turbo. It was tested on v4 and on Seed Audio.
• The row's audio Download, Recreate and Delete buttons.
• A real Seed Audio generation, the Japanese locale, and the mobile app.
• The sidebar layout beyond a quick look at 900 px, where the counter and quote behaved the same.