**Problem/Motivation** On a site where I have this enabled, the client noticed that the ElevenLabs-generated audio files were playing only the first 5,000 characters of text, even though the Elevenlabs limit is 10,000 characters.\ \ `ElevenlabsProvider::textToSpeech()` splits text longer than 5,000 characters into\ chunks, requests one MP3 per chunk, and then joins the results. The join is a raw\ byte append:\ \ <span dir="">`audio_binary_data = _NULL_;\ foreach (`</span>`audio_files as $audio_file) {$audio_binary_data .= $audio_file->getBinary();}`\ \ Every response from the ElevenLabs API is a complete MP3 file, not a bare audio\ stream. Each one begins with an ID3v2 tag and a Xing/Info header frame that declares\ the frame count and duration of that segment. Appending the files leaves the first\ segment's Xing header at the front of the result, so every player that reads it, \ including the HTML5 `<audio>` element, reports and plays only the first segment.\ \ The remaining audio is physically present in the file but unreachable. The symptom is\ that content longer than 5,000 characters is silently cut off at the first chunk\ boundary, with no error logged and a file whose byte size looks correct. ### Evidence Concatenating three real API responses the way the module does, then parsing the MPEG\ frames:\ \ `file size: 16295523`\ `ID3 tags found: 3`\ `at byte 0 (size 45)`\ `at byte 5431841 (size 45) <- stray ID3 tag mid-stream`\ `at byte 10863682 (size 45) <- stray ID3 tag mid-stream`\ `Xing/Info headers found: 3`\ `at byte 45: Info declares 12995 frames -> 339.5 s`\ `at byte 5431886: Info declares 12995 frames -> 339.5 s`\ `at byte 10863727: Info declares 12995 frames -> 339.5 s`\ \ `actual stream duration : 1018.5 s (17.0 min)`\ `duration players trust : 339.5 s (5.7 min) <- from the FIRST Info header`\ `hidden/unreachable : 679.0 s (67% of the audio)`\ \ \ **There are four further defects in the same block**\ \ These are smaller, but they live in the \~40 lines.\ \ **1.** Words are run together at chunk-rebuild boundaries. `wordwrap()` replaces the\ break character with `"\n", explode()` then discards it, and the rebuild loop\ concatenates the pieces with no separator. <span dir="">2. An empty chunk is POSTed when the first wrapped line reaches the limit. `_`</span>`chunks_rebuilt[0]` is seeded with `''`, and the guard is `strlen($updated) < `<span dir="">`chunk_length`. If the first line is exactly 5,000 characters, the guard is false on the very first iteration, so index 0 is left empty and the loop moves to index 1</span>. <span dir="">That empty string is sent to the API as a request.</span> <span dir="">3. A single token longer than the limit is sent unsplit. `wordwrap()` is called with `cut = FALSE`, so an unspaced run like a long URL, or a script that does not use spaces is never broken:</span> <span dir="">`input: one 12000-char token -> chunks: 2, sizes [0, 12000]`</span> <span dir="">The 12,000-character chunk exceeds the API limit and the request fails.</span> <span dir="">4. The final chunk is conditioned on itself. `previous_text` and `next_text` are added to </span>`configuration` inside the loop but never removed, so on the last iteration, \ where there is no next chunk, `next_text` still holds the value set during the previous\ iteration, which is the chunk being sent:\ \ `1.1.1 behaviour:`\ `request for CHUNK-A prev=(unset) next=CHUNK-B`\ `request for CHUNK-B prev=CHUNK-A next=CHUNK-C`\ `request for CHUNK-C prev=CHUNK-B next=CHUNK-C <- told its own text follows it`\ \ The model is told that the passage it is reading is also the passage that comes next. ## Steps to reproduce 1. Configure the ElevenLabs provider with any voice and \`eleven_multilingual_v2\`. 2. Call text-to-speech with a body of more than 5,000 characters (e.g. via an\ `ai_automators` text-to-audio field, or the AI API Explorer). 3. Play the resulting file in a browser.\ \ **Expected:** the whole text is read aloud.\ **Actual:** playback ends at the first chunk boundary. The file size reflects the full\ text; the reported duration does not. ## Proposed resolution Strip the per-segment container metadata before joining, and rewrite the splitter.\ \ **Joining.** Add `concatenateMp3()`, which for each segment removes leading ID3v2\ tag(s), a trailing ID3v1 tag, and a leading Xing/Info/VBRI header frame, then appends the\ bare MPEG frames. The result carries no misleading declared duration. ElevenLabs returns CBR MP3 (`mp3_44100_128` by default), so players derive the correct duration from bitrate and file size.\ \ **Splitting.** Replace `wordwrap()` + the rebuild loop with `splitText()`, which splits\ on whitespace via `preg_split('/(\s+)/u', ...)` with `PREG_SPLIT_DELIM_CAPTURE` so the separators are preserved, counts length with `mb_strlen()`, never emits an empty chunk, and falls back to `mb_str_split()` for a single token longer than the limit, on\ character boundaries, so multibyte text is not cut mid-character (`wordwrap()` with\ `cut = TRUE` would corrupt it).\ \ **Conditioning.** `previous_text` and `next_text` are unset at the top of each iteration\ and set only when a neighboring chunk actually exists, so the final request no longer\ carries a stale `next_text`.\ \ **Empty input.** When the text contains nothing to read, `splitText()` returns an empty\ array and the method now throws `AiBadRequestException('No text to convert to speech.')` rather than sending a request. Previously, this surfaced as `'No audio found'` after a wasted API call. ## User interface changes None. ## API changes No signature changes. Four \`protected\` methods are added: `splitText()`,\ `concatenateMp3()`, `stripMp3Metadata()` and the `mpegFrameLength()` helper.\ \ Two behavioral notes: - `TextToSpeechOutput`'s raw-output argument now carries the joined binary rather than\ the last chunk's response, which is what callers would reasonably expect.\\ - Text with nothing readable in it now throws `AiBadRequestException` before any request is made, instead of after one. ## Data model changes None. ## AI note AI (Claude) was used to help with this analysis, code and documentation, but a real human (me) validated it end-to-end on a real project.