Bespoken 5.3.0: Long-Form Audio, Credit Tracking, and Multisite Support

When I shipped Bespo­ken 5.2.0 ear­li­er this month, the focus was on refin­ing pro­nun­ci­a­tion rules, mak­ing it eas­i­er to strip unwant­ed char­ac­ters from nar­ra­tion scripts and giv­ing you visu­al feed­back in the set­tings UI. That release laid impor­tant ground­work, but I knew the plu­g­in still had a hard ceil­ing: if your entry text was too long, Eleven­Labs would reject it out­right. And if you were run­ning a mul­ti­site install, things could break in sub­tle, frus­trat­ing ways. Sor­ry about that!

This next release tack­les both of those prob­lems head-on, along with a long list of improve­ments to the audio gen­er­a­tion expe­ri­ence. Here’s what’s new.

Text Chunking and Audio Concatenation

This is the head­line fea­ture. Bespo­ken now auto­mat­i­cal­ly splits long text into chunks at para­graph and sen­tence bound­aries before send­ing each piece to Eleven­Labs. When all chunks are gen­er­at­ed, they’re con­cate­nat­ed into a sin­gle MP3.

Pre­vi­ous­ly, if your entry text exceed­ed the mod­el’s char­ac­ter lim­it, the API call failed and told you that the text was too long. Now, Bespo­ken han­dles it trans­par­ent­ly — you hit ​“Gen­er­ate” and get back one seam­less audio file, regard­less of text length.

The chunk­ing log­ic pre­serves para­graph bound­aries through both the fron­tend and back­end pipelines, so splits hap­pen at nat­ur­al points in your con­tent rather than mid-sen­tence.

Request Stitching for Natural Transitions

Chunk­ing alone would pro­duce audi­ble seams between seg­ments — shifts in pac­ing, tone, or inflec­tion at chunk bound­aries. To solve this, Bespo­ken uses Eleven­Labs’ request stitch­ing API. Each chunk is con­di­tioned on the sur­round­ing text, so the mod­el main­tains con­sis­tent prosody across the full nar­ra­tion.

Request stitch­ing is auto­mat­i­cal­ly dis­abled for eleven_v3 (which does­n’t sup­port it, at least not yet) and for sin­gle-chunk gen­er­a­tions where it’s unnec­es­sary.

ElevenLabs Credit Display and Cost Estimates

The Bespo­ken field now shows your remain­ing Eleven­Labs cred­its, the reset date, and a usage bar — pulled from the Eleven­Labs sub­scrip­tion API when the page loads. Before gen­er­at­ing audio, you can see at a glance whether you have enough cred­its.

Along­side this, the plu­g­in now cal­cu­lates an esti­mat­ed cred­it cost based on your text length and the select­ed voice’s mod­el. Stan­dard mod­els (v3, mul­ti­lin­gual) use a 1x mul­ti­pli­er, while tur­bo and flash mod­els cost 0.5x. If the esti­mate exceeds your remain­ing cred­its, you’ll see a warn­ing before you com­mit to the gen­er­a­tion.

Progress and Monitoring Improvements

With audio gen­er­a­tion now poten­tial­ly span­ning mul­ti­ple chunks and tak­ing longer to com­plete, the progress UI need­ed to keep up.

Chunk Progress Reporting

The progress indi­ca­tor now shows which chunk is being gen­er­at­ed — for exam­ple, ​“Gen­er­at­ing audio: chunk 3 of 7.” This gives you a clear sense of how far along a mul­ti-chunk job is.

Complete Message History

All progress mes­sages are now accu­mu­lat­ed in the data­base. The fron­tend replays any unseen mes­sages on each poll, so the full mes­sage his­to­ry is always vis­i­ble — even when chunks process faster than the 1‑second polling inter­val. You can expand the progress com­po­nent to see every step of the gen­er­a­tion pipeline.

Smarter Timeouts

The fron­tend job mon­i­tor now uses a stall-based time­out instead of a fixed poll count. It waits for 3 min­utes of no progress change before report­ing a time­out, so long mul­ti-chunk jobs don’t false­ly time out while they’re active­ly work­ing. On the back­end, the queue job’s time-to-reserve now scales with text length and chunk count, pre­vent­ing the queue run­ner from killing jobs that legit­i­mate­ly need more time.

Stale Job Detection

If a job gets stuck in ​“run­ning” sta­tus for more than 10 min­utes (e.g., the queue work­er was killed), it’s auto­mat­i­cal­ly marked as failed when you view the gen­er­a­tion his­to­ry. No more phan­tom ​“run­ning” indi­ca­tors that nev­er resolve.

Multisite Support

This was a sig­nif­i­cant bug fix. On mul­ti­site Craft installs, gen­er­at­ing audio on a non-pri­ma­ry site would fail with an ​“Ele­ment not found” error because the con­troller was­n’t resolv­ing the cor­rect site con­text. This release threads site con­text through the entire pipeline:

• Action URLs now explic­it­ly include the site han­dle para­me­ter
• The gen­er­a­tion his­to­ry modal fil­ters by the cur­rent site
• The queue job car­ries the orig­i­nat­ing site ID

If you’re run­ning Bespo­ken on a mul­ti­site install, this update should resolve the issues you’ve been hit­ting.

Upgrading

This update will include a data­base migra­tion. Run craft migrate after updat­ing to apply it. The migra­tion is non-destruc­tive — it adds columns to the exist­ing bespoken_audiogenerations table.

If you’re using the queue for audio gen­er­a­tion (and you should be), make sure your queue work­er is run­ning: craft queue/listen.

TLDR: Bespo­ken now han­dles long-form con­tent, shows you what it’ll cost before you gen­er­ate, works cor­rect­ly on mul­ti­site installs, and gives you much bet­ter vis­i­bil­i­ty into what’s hap­pen­ing dur­ing gen­er­a­tion. Give it a try and let me know how it goes.