Paxa Labs

Synthesize a short line with Paxa TTS and play it through this machine's speakers. Blocks until playback finishes. Waits for the currently playing audio but jumps ahead of queued long-form segments. For long content use queue_speech. Costs 10 credits per 1000 characters, 12.5 in a designed voice (voice_...). Direction tags (paxa-tts-flash-v2, every voice): words in square brackets written into the text set the delivery from that point on, in English, Thai, or Chinese, e.g. [whispering], [speaking slowly, close to the microphone], [sad:0.5], or a blend [happy:0.7+sad:0.3]; [neutral] returns to the voice's own delivery. Sounds such as [laughs], [sighs], [gasps] play where written; [um] and [hmm] are spoken; <emph>...</emph> and <whisper>...</whisper> mark a span. [short pause], [pause], and [long pause] add silence after the word before them, and [pause 300ms] an exact length up to 500 ms. Nothing in brackets is spoken, tag characters bill as text, and a tag the model cannot do is refused before any charge. On paxa-tts-flash-v1 only khanomchan and kaprao read tags, from a fixed emotion list, and no voice reads pause tags.

Read long content aloud (stories, articles, books). Splits the text into segments, synthesizes ahead while playing, and returns at once with the segment count and the estimated credits; segments play in order after anything already queued. There is no handle: control_playback monitors and controls the whole queue. Use speak instead for one short line that should play now and wait, and text_to_speech for a file without playback. Costs 10 credits per 1000 characters, 12.5 in a designed voice (voice_...). Direction tags (paxa-tts-flash-v2, every voice): words in square brackets written into the text set the delivery from that point on, in English, Thai, or Chinese, e.g. [whispering], [speaking slowly, close to the microphone], [sad:0.5], or a blend [happy:0.7+sad:0.3]; [neutral] returns to the voice's own delivery. Sounds such as [laughs], [sighs], [gasps] play where written; [um] and [hmm] are spoken; <emph>...</emph> and <whisper>...</whisper> mark a span. [short pause], [pause], and [long pause] add silence after the word before them, and [pause 300ms] an exact length up to 500 ms. Nothing in brackets is spoken, tag characters bill as text, and a tag the model cannot do is refused before any charge. On paxa-tts-flash-v1 only khanomchan and kaprao read tags, from a fixed emotion list, and no voice reads pause tags.

Inspect and control the shared audio queue that speak, queue_speech, play_audio, and the voice design tools play through. status reports what is playing, what is queued, and any synthesis failures. pause holds playback until resume or clear, and audio queued meanwhile waits; pause is not supported on Windows, where skip or clear stops playback instead. skip drops the current segment and is refused when nothing is playing. clear stops playback and discards every queued segment, which must be queued again to be heard; audio files already written stay on disk. Free (no API call).

Play a local audio file through this machine's speakers, for example a file produced earlier by text_to_speech. Waits for the audio playing now, plays ahead of queued long-form speech, and blocks until the file ends; control_playback skip or clear stops it early. Fails without playing when the file is missing, is not a playable format, or no audio player is installed; raw pcm, mulaw, and alaw files have no header and cannot be played. To say new text, use speak instead. Free (no API call).

Synthesize speech with Paxa TTS and save it as an audio file without playing it: voiceovers, narration, audiobooks, phone prompts. To hear speech instead, use speak (a short line) or queue_speech (long reads); play_audio plays a saved file later. The file goes to output_path, replacing a file already there, or to a new timestamped name in PAXA_OUTPUT_DIR. For several speakers in one file, use dialogue_to_speech. Text beyond 20000 characters is split at paragraphs and sentences and the parts are joined into one file (mp3, opus, and wav need ffmpeg for that). With subtitles set, an srt or vtt caption file that matches the voice word for word is written next to the audio. Costs 10 credits per 1000 characters, 12.5 with subtitles, 12.5 in a designed voice (voice_...), 15 with both; every format and sample rate costs the same. Direction tags (paxa-tts-flash-v2, every voice): words in square brackets written into the text set the delivery from that point on, in English, Thai, or Chinese, e.g. [whispering], [speaking slowly, close to the microphone], [sad:0.5], or a blend [happy:0.7+sad:0.3]; [neutral] returns to the voice's own delivery. Sounds such as [laughs], [sighs], [gasps] play where written; [um] and [hmm] are spoken; <emph>...</emph> and <whisper>...</whisper> mark a span. [short pause], [pause], and [long pause] add silence after the word before them, and [pause 300ms] an exact length up to 500 ms. Nothing in brackets is spoken, tag characters bill as text, and a tag the model cannot do is refused before any charge. On paxa-tts-flash-v1 only khanomchan and kaprao read tags, from a fixed emotion list, and no voice reads pause tags.

Synthesize a script of lines, each in its own voice, into one audio file without playing it: podcasts, interviews, radio plays, lessons and phone flows with more than one speaker. Pass the lines in speaking order with a voice each; gap_ms of silence goes between lines. For a single voice use text_to_speech; to hear a conversation now, call queue_speech once per line in that line's voice; play_audio plays the saved file. The file goes to output_path, replacing a file already there, or to a new timestamped name in PAXA_OUTPUT_DIR. A script over 100 lines or 20000 characters is sent in parts split between lines and joined with the same gap (mp3, opus, and wav need ffmpeg for that). Returns each line's start and end in the audio. With subtitles set, an srt or vtt caption file is written next to the audio, and no cue spans two lines. If synthesis fails partway, the lines already delivered whole, which were paid for, are saved and the result names the line to resume from. Costs the same as text_to_speech for the characters of every line: 10 credits per 1000, 12.5 with subtitles, 12.5 for a line in a designed voice (voice_...), 15 with both; the gap is free. Direction tags (paxa-tts-flash-v2, every voice): words in square brackets written into the text set the delivery from that point on, in English, Thai, or Chinese, e.g. [whispering], [speaking slowly, close to the microphone], [sad:0.5], or a blend [happy:0.7+sad:0.3]; [neutral] returns to the voice's own delivery. Sounds such as [laughs], [sighs], [gasps] play where written; [um] and [hmm] are spoken; <emph>...</emph> and <whisper>...</whisper> mark a span. [short pause], [pause], and [long pause] add silence after the word before them, and [pause 300ms] an exact length up to 500 ms. Nothing in brackets is spoken, tag characters bill as text, and a tag the model cannot do is refused before any charge. On paxa-tts-flash-v1 only khanomchan and kaprao read tags, from a fixed emotion list, and no voice reads pause tags.

Make a new voice from a written description and play the line it speaks through this machine's speakers. Use it when the user wants a voice the catalog does not have. Describe who is speaking and how they sound in one or two sentences: age, gender, role, pitch and texture, pace and manner, where it will be heard, and the language it speaks (a designed voice speaks one language: the one the description names, otherwise the one it is written in), such as "A calm Bangkok pharmacist in her thirties who explains dosages slowly and kindly, speaking Thai". Leave out real people's names, sound effects, and what the voice should say. The result is a draft, not yet a voice: refine_voice changes it, save_voice keeps it. Costs 5 credits per call, the line included; a retry designs again and is charged again.

Change a draft voice in words and play its new line. The instruction says what to change, such as "a little lower and slower" or "warmer, and older"; the rest of the voice stays, so change one quality at a time. Each call returns a new draft handle and the earlier handles keep their versions, so the user can go back to one. save_voice keeps the draft the user picks. Refining never changes a voice already saved: saving a refined draft makes another voice, and rename_voice changes only a name. Costs 5 credits per call, the line included; a retry refines again and is charged again.

Save a draft from design_voice or refine_voice as one of the account's voices, and get its id (voice_...), which speak, queue_speech, and text_to_speech accept as voice on paxa-tts-flash-v2, for this account only. Save only the draft the user picked. Saving the same draft again returns the same voice and uses no slot. A saved voice's sound never changes: design a new voice for a different sound. The plan sets how many voices the account keeps (Free 1, Starter 3, Standard 10, Pro 50); list_voices shows the slots in use. Free.

Rename one of the account's designed voices, given its voice_ id from list_voices or save_voice. The id and the sound stay the same, so anything that uses the id keeps working. The new name takes 1 to 60 characters, and a catalog voice's name is refused. To change how a voice sounds, refine a draft and save it as a new voice; to remove one, use delete_voice. Free.

Delete one of the account's designed voices and free its slot. The id is the voice_ id list_voices shows; it is looked up first, so an id the account does not have is answered without deleting anything, and catalog voices cannot be deleted. Cannot be undone: every later request in the voice fails, and designing again does not bring the same sound back. Use it only when the user asks to delete that voice; to change only its name, use rename_voice. Free.

Translate text into any of 15 languages, from any of them, with Paxa Translate: Thai (th), English (en), Chinese (zh), Japanese (ja), Korean (ko), Arabic (ar), French (fr), German (de), Hindi (hi), Indonesian (id), Malay (ms), Portuguese (pt), Russian (ru), Spanish (es), Vietnamese (vi). The source language is detected per segment, so mixed-language text works; a segment already in the target language comes back unchanged. Accepts one string or up to 200 segments that share context, and returns each translation (numbered for segments) with its detected source language, any review flag, and the credits used. For text in a PDF or image, run ocr_document first. Costs 25 credits per 1000 text characters per rendering (minimum 2 credits), so 1000 characters cost 25 credits, or 50 with alternatives 2; instructions, context, examples, do_not_translate strings, and glossary terms bill at 8 per 1000.

Run Paxa OCR on a local PDF, PNG, JPEG, or WebP file and return its content as GitHub-flavored Markdown (or structured blocks). Reads Thai, English, Simplified Chinese, and other languages as printed, pages that mix them included, with no language setting; text set in two columns reads column by column. output "structured" returns typed blocks as JSON instead of Markdown. Use extract_fields instead for specific values (a total, a date, an ID number) checked and typed against a schema. Up to 50 pages and 10 MiB per file. Costs 6.5 credits per page.

Extract typed fields from a local PDF, PNG, JPEG, or WebP document (up to 20 pages and 10 MiB) with Paxa Document Extraction. You write a schema of fields; every value returned is text printed in the document, read as the field's type, or null with the reason. Nothing is inferred or looked up. Built for Thai documents (receipts, tax invoices, forms, IDs) and fine on English ones. Use ocr_document instead for a document's full text. Returns a status line, the fields as JSON (null where a value is not printed), then the missing, unverified, truncated, and assumed fields. Costs 13 credits per page for a schema of up to 50 leaf fields and 19.5 per page for 51 to 100; the page reading and the evidence spans cost nothing extra. A schema the API refuses costs nothing and names the field and reason. The schema parameter summarizes the dialect; https://paxalabs.com/docs/extraction has every type and key.

Transcribe a local recording or video of speech in Thai, English, Mandarin Chinese, or another language, code-switching included (mp3, wav including mu-law and A-law phone recordings, flac, ogg, m4a, aac, webm; or mp4, mov, mkv, avi and other video files, whose sound track is extracted locally with ffmpeg; up to 120 minutes and 95 MiB) with Paxa STT. The transcript comes back inline, so nothing needs to be read from disk afterwards; optionally it is also saved as txt, json (with word timings and segments), srt, or vtt, named after the recording. Existing files are never overwritten. Detects the language and can label speakers; nothing is translated. For one live answer from the microphone, use listen instead. Terms configured in PAXA_VOCABULARY or PAXA_VOCABULARY_FILE are pinned on every call. Costs 8.33 credits per minute of audio (0.1 minimum); subtitles and word timings cost nothing extra. Long recordings take minutes to process.

Record one spoken turn from this machine's microphone and return its transcript (Thai, English, Mandarin Chinese, or another language, code-switching included). Paxa's realtime transcription detects when the user starts and stops talking, so the call returns as soon as the turn ends; the audio goes straight to the API and nothing is written to disk. Waits for any speech this server is playing to finish first. Use it after speak when you asked the user a question and want their spoken answer; for a recording already on disk, use transcribe_audio instead. Returns an empty transcript when nobody spoke within wait_seconds. Costs 12.5 credits per minute of audio sent, silence included (a few seconds of listening costs well under one credit). Needs a microphone recorder on this machine (ffmpeg or sox; parecord or arecord on Linux) and Node 22 or newer; the host app must have microphone permission.

Report a wrong or poor output, an API problem, or a platform issue to the Paxa Labs team, tied to the request ids the other tools print ("Request id: ..."). A rating alone on at least one request id is a complete report. For a wrong output, add an example with the exact input, the output received, and the output expected. Returns the report's id once it is filed. Free, no credits. Use it when the user says an output was wrong, or asks to report a problem; do not file reports the user did not ask for.

List the Paxa TTS voices in one response (no paging): recommended voices first, the account's designed voices with their slots, then the catalog roster per model with ids, names, gender, language, and character notes. Call it when the user asks what voices exist or wants a different one, and before save_voice to see free slots; pass an id from it as voice to speak, queue_speech, or text_to_speech. The default voice needs no lookup. Free.

無需輸入參數。

List the available Paxa models (speech, transcription, realtime transcription, translation, OCR, and extraction) with their limits and credit pricing, and the languages translate writes. Call it to check a limit or a price before a large job; voices are in list_voices, and the credit balance in get_account. Free.

無需輸入參數。

Connect this machine to the user's Paxa account when no API key is configured: opens the Paxa sign-in page in the user's browser with a code filled in, and the user approves there (new accounts include free credits). Every Paxa tool then works, in every session on this machine, until the user signs out or leaves it unused for 30 days. Waits up to wait_seconds for the approval; if it has not come yet, call sign_in again to keep waiting on the same code. When no browser opens, it returns the page and the code at once for you to give the user. Answers at once when already signed in or when PAXA_API_KEY is set. Free.

Show the account behind the API key or sign-in, read from the API on each call: credit balance, spending cap, plan, and rate limits. Call it before a large paid job, or after a call fails for credits or a rate limit; prices per model are in list_models. Free.

無需輸入參數。