How do I give an agent a voice?

Q: How do I give an agent a voice?

speak it, read it back, and check the transcript say out loud with elevenlabs on monid. it returns a hosted link rather than audio bytes, so download the file in the same run: the link is documented to last an hour and the file a week what comes back: a download link on Monid's own file service, never inline bytes, with a documented one-hour link expiry and a seven-day file expiry asking again is safe: fetching the run a second time mints a fresh link and the old one keeps working, so a link that has expired costs nothing to replace but the file clock does not restart: the seven days run from the original call, not from the refetch, so a fresh link on day seven buys you hours rather than another week and there IS a poll: an earlier version of this page said there is no polling and no job id anywhere here, which contradicts the platform's own documented run-then-poll loop; fetching the run by id IS that poll one field halves it, and you must pass it: naming the faster model cut the bill exactly in half on the same words on both runs, and it produced slightly more audio, so the per-minute gap is a little wider than 2x. Left unset, the default is the dearer model, so the saving is something you ask for rather than something you get the default voice is not in the list: the documented default voice id does not appear among the voices the voice endpoint returns, so pick one from the list rather than relying on the default. The list is account-scoped and came back with 42 and then 44 entries, so read it rather than quoting a count

Elevenlabs APIs

Turn text into speech

Say this to your agent

>say <text> out loud with elevenlabs on monid. it returns a hosted link rather than audio bytes, so download the file in the same run: the link is documented to last an hour and the file a week

What the agent runsdone

elevenlabs /v1/text-to-speech$0.05 / 1k chars

elevenlabs /v1/text-to-dialogue$0.10 / 1k chars

elevenlabs /v1/voicesread-only

minimax /v1/t2a_v2catalog rate

What came backone sentence, round trip · 2026-09-25

The same sentence, two modelsUS dollars for 69 characters

the default model$0.0069

the faster model$0.00345

what comes backa download link on Monid's own file service, never inline bytes, with a documented one-hour link expiry and a seven-day file expiry

asking again is safefetching the run a second time mints a fresh link and the old one keeps working, so a link that has expired costs nothing to replace

but the file clock does not restartthe seven days run from the original call, not from the refetch, so a fresh link on day seven buys you hours rather than another week

and there IS a pollan earlier version of this page said there is no polling and no job id anywhere here, which contradicts the platform's own documented run-then-poll loop; fetching the run by id IS that poll

one field halves it, and you must pass itnaming the faster model cut the bill exactly in half on the same words on both runs, and it produced slightly more audio, so the per-minute gap is a little wider than 2x. Left unset, the default is the dearer model, so the saving is something you ask for rather than something you get

the default voice is not in the listthe documented default voice id does not appear among the voices the voice endpoint returns, so pick one from the list rather than relying on the default. The list is account-scoped and came back with 42 and then 44 entries, so read it rather than quoting a count

$0.006969 characters spoken

Read it back and check it

Say this to your agent

>transcribe the audio you just made with elevenlabs on monid, then diff the transcript against the text you sent character by character. expect a word with a diacritic or an unusual spelling to come back wrong and expect the final full stop to be missing. try that word in keyterms, but check the result: on our second run keyterms turned a wrong word into a non-word rather than fixing it

What the agent runsdone

elevenlabs /v1/speech-to-text$0.22 / hour

elevenlabs /v1/forced-alignment$0.22 / hour

elevenlabs /v1/text-to-speech$0.05 / 1k chars

What came backone sentence, round trip · 2026-09-25

it came back wrongon a clean machine-generated clip with no noise, no accent and no crosstalk, a word came back as a different word and the closing full stop was dropped; that dropped full stop reproduced on a second sentence in all three runs

what breaks is not always the brand nameon a second sentence the brand name and the unusual surname both survived and the word with a diacritic did not: Wroclaw came back as Rockwool

and keyterms does not reliably fix itthe first run it did. On the second it made it worse, turning the wrong word into a non-word, at the same $0.000069 extra. So treat keyterms as worth a try and then check, not as the fix

what that impliesif a synthetic clip mis-transcribes a proper noun, a real recording will too; tell the transcriber the names it should expect rather than checking afterwards

the round trip is cheaptext out loud and back into text cost $0.003756 in total, which is the cheapest verification step on this page

$0.003756the whole loop

Making audio against reading it

Say this to your agent

>before you plan a job, measure what a minute of audio costs to make and what a minute costs to read, on the models you will actually use: characters per minute is not a constant, and the model this command picks by default is the expensive one. then multiply both out to a thousand minutes

What the agent runsdone

elevenlabs /v1/text-to-speech$0.05 / 1k chars

elevenlabs /v1/speech-to-text$0.22 / hour

elevenlabs /v1/music$0.15 / minute

What came backone sentence, round trip · 2026-09-25

One minute of audioUS dollars a minute

compose music$0.15

speak it$0.0469

transcribe it$0.003667

the multiplereading is roughly twelve to forty times cheaper than making, depending which model and which route; on defaults it came out at about 27x

the rate is not a constant, so measure itcharacters per minute of delivered audio came out at 938 on the first sentence and at 869 and 976 on a second, for the two models, so the same text buys a different number of minutes depending on the voice

and the default model is the dear onethe halving only happens if you pass the model explicitly; left to the default, a thousand minutes of speech is about $97 rather than about $43

music has two shapes, not oneone provider bills per minute and the other bills a flat price per song regardless of length, which came out at about $57 a thousand minutes against about $150, so the cheaper composer depends entirely on how long your pieces are

per character ranks differentlythe two-voice route costs the same per character as the single-voice one and less per minute, because two speakers with a laugh in them talk more slowly

$0.00arithmetic on measured rates

Music and a sound bed

Say this to your agent

>make me <n> seconds of <mood> music with elevenlabs on monid, and a <n> second sound bed. this provider bills exactly the duration you ask for, so ask for exactly what you need. if you use the other music provider instead, note that it takes no duration parameter at all and bills a flat price per song, so short pieces are dear there and long ones are cheap

What the agent runsdone

elevenlabs /v1/music$0.15 / minute

elevenlabs /v1/sound-generation$0.12 / minute

minimax /v1/music_generationcatalog rate

What came backone sentence, round trip · 2026-09-25

it is deterministicyou set the duration and it bills exactly that duration: $0.025 for ten seconds of music and $0.01 for a five-second bed, with no rounding surprise

so budget in secondsunlike the character-priced routes, there is nothing to measure afterwards; the number you ask for is the number you pay

the second vendor works and prices differentlyan earlier version of this page wrote it off as dead. On a second run it answered and billed a flat fifteen cents for a song of about two and a half minutes, with no duration parameter to set

a third one is the dead onea different provider's music endpoint, flagged unknown in the catalog rather than stable, timed out with a 504; it billed nothing, as every genuine failure here did

so ask for exactly what you need is provider-specificit is true of the per-minute provider and meaningless on the flat-rate one

$0.035ten seconds and five

One file, four different durations

Say this to your agent

>if you send the same audio to more than one endpoint, record what duration each one billed: on our runs one identical file was billed at four different durations by four endpoints on the same provider, spreading about fifty percent. one of them also enforces a 4.6 second minimum that is in no schema, and the CLI hides that rejection behind a generic 400 with no output file, so read the run record rather than the client error

What the agent runsdone

elevenlabs /v1/speech-to-text$0.22 / hour

elevenlabs /v1/forced-alignment$0.22 / hour

elevenlabs /v1/speech-to-speech$0.12 / minute

elevenlabs /v1/audio-isolation$0.12 / minute

What came backone sentence, round trip · 2026-09-25

The same 4.133 second file, billedseconds charged

transcription6 seconds

forced alignment5 seconds

isolation4 seconds

voice conversion4 seconds

the spreadone file, four endpoints on the same provider, four different billed durations: 67% apart on the first file and about 50% on a second, so the shape holds and the percentage does not

and the floor contradicts itselfthe endpoint that refuses a 4.55 second file for being under its 4.6 second minimum billed a 5.108 second file as four seconds, which is below the minimum it just enforced

why it matters at scalea pipeline that estimates cost from the file's real duration will be wrong in three different directions depending on which endpoint it used

the undocumented floorone endpoint enforces a 4.6-second minimum that appears nowhere in its schema, and rejects anything shorter

and the CLI hid itthat rejection printed only a generic HTTP 400 and wrote no output file, while still creating three real runs whose records carried the actual too-short message

$0.00the spread, not the cost

What this aisle cannot do

Say this to your agent

>tell me plainly what is not here before i design around it: there is no dubbing, and nothing in these three aisles streams a generated file back to me. do NOT conclude from that that live voice is impossible: a live phone call is metered by the second elsewhere on the platform, and the run stays open for the length of the call. check whether a cloned voice can be USED before you tell me one cannot be MADE: they are different questions

What the agent runsdone

elevenlabs /v1/voicesread-only

elevenlabs /v1/speech-to-speech$0.12 / minute

minimax /v1/t2a_v2catalog rate

What came backone sentence, round trip · 2026-09-25

no voice CREATION, which is not the same as no cloned voicesnothing in these aisles creates or clones a voice. But one provider's own voice setting states that system, cloned and AI-designed voices are all accepted, and it carries a timbre control that blends four voices into a new one, so a cloned voice made elsewhere can be used here

and the voice list is not a fixed 42it returned 44 on a second run and it is scoped to the account, so it is a list to read rather than a number to quote

no streaming, and that does NOT rule out live voiceno endpoint in these three aisles streams a generated file back to you; every route is request, wait, download. An earlier version of this page concluded from that that a real-time conversational agent is impossible. It is not: a phone call elsewhere on the platform is metered per second at about thirty cents a minute, the run stays open for the length of the call and stopping the run hangs up

a stream flag exists but is pinned offtwo endpoints do declare a stream parameter, fixed at false, so an agent grepping the schemas will find one and must read the constant rather than the key

no dubbingthere is a voice-to-voice conversion endpoint, and it is not the same thing as translating and re-voicing a track

read the notes, not the summaryfive of nine summaries describe the output as base64 audio and all five returned a hosted link instead; the notes block was the accurate one, and it documents base64 as a fallback for when the file cannot be saved rather than as something that never happens

$0.00know the holes first

Elevenlabs

How-to guide

Set up Monid.

Pick what you need. Say it in one sentence.

Turn text into speech

Read it back and check it

Making audio against reading it

Music and a sound bed

One file, four different durations

What this aisle cannot do

Read more

MiniMax H3 Max on Monid

LLM gateway or MCP gateway

Extract page content for RAG