Skip to main content

Chatterbox API

Uncensored — This model declares no content filter. SpicyAPI forwards the request without adding a platform content scan; legal and consent duties remain with the customer.Browse every model at this filtering level
resemble-ai/chatterbox-v1/text-to-speech

Resemble AI's open-source Chatterbox text-to-speech reads up to 1,000 characters of English into a 24 kHz mono WAV, in its built-in voice or the voice of a short reference clip you send. Expressiveness and variation are adjustable.

From $0.036/request
Run model

Workbench

Input

Text to speech
119/1000
Up to 1,000 characters of English, read exactly as written. Tags such as <laugh> are read aloud; spell sounds out instead. at least 1 characters · at most 1000 characters

This model reads tags like [sighs] aloud. Spell sounds out instead, like “Mmm…” or “Haha”.

$0.036 flat per request, up to 1,000 characters

0/1

Drop an audio file here, or browse · 1 file · Max 90 MB · MP3 · WAV

0 / 1 files
Optional MP3 or WAV clip of the voice to imitate: upload one or give a public URL. The reading takes on that clip's timbre; the words in the clip are not repeated. A few seconds of one clean voice without music works best. Use only voices you have the right to use. Does not change the price. at least 1 characters
Expressiveness
How emotional the reading is, from 0 (flat) to 1 (dramatic). Defaults to 0.5. no lower than 0 · no higher than 1
Temperature
Variation between takes, from 0.05 to 2. Defaults to 0.7. no lower than 0.05 · no higher than 2
How closely the reading follows the text's natural pacing, from 0.1 to 1. Defaults to 0.5; lower it to slow a rushed reading. no lower than 0.1 · no higher than 1
Safety checkerOff

SpicyAPI does not run a safety checker on this model. Your prompt and media go to the model as written, and the model itself decides what comes back.

Checking your account before enabling submission and pricing.

Output

Ready for a new run
Example result
0:00 / –:––
Avatar
Speech to text
Lip sync

Related models and next tasks

Missing the model you need?

Tell us what to add and where you would use it. Requests go to the people who decide what gets onboarded next.

One built-in voice, three sliders, and no tags to fight

Chatterbox is Resemble AI's open-source text-to-speech model: send up to 1,000 characters of English and it reads them back in its one built-in voice, as a WAV file at 24 kHz. There is no voice list and no cloning on this endpoint — exaggeration, temperature and guidance_scale change how that single voice performs, not who it is. Every request costs the same flat price regardless of length, and bracketed or angle-bracket tags such as <laugh> are read aloud as words rather than performed, so sounds need to be spelled out. SpicyAPI applies no content filter, and in our tests our strongest test scripts were read in full.

Open the playground
Voice
One built-in English voice; no list, no cloning
Characters per request
Up to 1,000, English only
Output
24 kHz, 16-bit mono WAV
Pricing
Flat $0.036 per request, any length up to the limit
Tags
<laugh> and similar are read aloud, not performed
Licence
Open source (MIT), Resemble AI, released May 2025
Voice
One voice, tuned three waysThere is no voice_id field because there is only one voice. exaggeration, temperature and guidance_scale reshape how it performs a line — calmer or more dramatic, more or less varied between takes — without ever changing which voice is speaking.
Pricing
A flat price, independent of lengthA six-character line and a 1,000-character paragraph cost the same. That makes short character lines comparatively expensive per character and long-form narration comparatively cheap — plan a batch of short lines accordingly.
Tags
Spell the sound out, don't tag itIn our tests, bracketed cues such as [sigh] or angle-bracket tags such as <laugh> were read aloud as literal words rather than performed as sounds. A spelled-out 'mmm...' or 'ha...' works where a tag does not, and raising exaggeration pushes a spelled-out sound further toward dramatic.

What the Chatterbox API does well

  • A voice with no setup decision to make

    There is nothing to pick: every request uses the same built-in voice, so there is no casting step, no comparing presets and no risk of a series drifting between voices because someone picked a different one halfway through.

  • Three sliders instead of a prompt to write

    exaggeration sets how emotional the reading is, temperature adds variation between takes, and guidance_scale controls how closely the delivery follows the text's natural pacing. Together they cover a wider range of deliveries than a single voice usually offers, without needing a text description of tone.

  • A flat price that is easy to budget

    Every request costs the same amount, whatever the character count up to the 1,000-character limit. A backlog of lines is a straight multiplication — line count times the flat price — with no need to estimate total characters first.

  • Lossless WAV output

    The result is a 24 kHz, 16-bit mono PCM WAV file rather than a compressed format, which matters when the clip is going to be mixed, layered with effects, or re-encoded downstream rather than played back as-is.

  • An open-source model behind a hosted endpoint

    Chatterbox is released under the MIT licence, but running it well still means managing GPU inference. This endpoint removes that step: the same open weights, called over HTTP, on a flat per-request price.

  • Unfiltered scripts read as written

    SpicyAPI adds no content filter on top of the model. In our tests, Chatterbox read our strongest first-person and narrated test scripts in full, including spelled-out sounds, without skipping or cutting the clip short.

Who Chatterbox is for

  • Game and animation studios

    Short character lines where energy matters more than timbre choice

    One voice with an exaggeration dial covers a lot of ground for bark lines, shouts and quick reactions, and the flat per-request price makes a large batch of short lines predictable to budget rather than punishing per character.

  • Prototyping teams

    A placeholder voice before the real casting decision

    Put a voice on a demo, a storyboard or a chatbot flow before committing to a final voice elsewhere in the catalog. The flat price per line keeps early iteration cheap regardless of how many takes a script goes through.

  • Uncensored audio fiction

    Narration that keeps going to the end of the scene

    In our tests, our strongest monologue and narration scripts were read in full, and spelled-out sounds were voiced rather than skipped. For a fiction catalog built around that material, a narration step that does not flinch is what keeps the pipeline usable.

  • Video and short-form creators

    A voice track to drop onto a timeline

    Generate the line as a WAV file, then hand it to a talking-avatar or lip-sync model, or cut it straight into an edit. The lossless output holds up through a mix pass better than a compressed file would.

How to call the Chatterbox API

  1. Write the script in English, under 1,000 characters

    Punctuate for the pauses you want. Longer scripts need to be split at sentence boundaries into several requests and stitched together afterward; each request is billed at the flat rate regardless of length.

  2. Spell out sounds instead of tagging them

    Tags such as <laugh> or [sigh] are read aloud as words in our tests, not performed. Write 'ha...', 'mmm...' or 'ahh...' directly in the text and raise exaggeration for a more dramatic reading of it.

  3. Set exaggeration for the emotional range

    0.5 is the default and sounds natural for most narration. Raise it toward 1 for dramatic, breathless or shouted lines; lower it toward 0 for flat, calm delivery.

  4. Adjust guidance_scale if the pacing feels off

    guidance_scale defaults to 0.5. Lowering it toward 0.3 slows down a reading that comes out too fast or too dramatic; raising temperature adds more variation between separate takes of the same line.

  5. Read the WAV off the task result

    The task returns one 24 kHz, 16-bit mono WAV file, kept in SpicyAPI storage for 14 days. A failed request is not charged.

Chatterbox worked examples

  • A dramatic character shout

    Raising exaggeration pushes a short line toward a more dramatic, breathless delivery -- useful for a game bark or an animated reaction.

    {"text": "Wait -- don't you dare walk out that door!", "exaggeration": 0.9, "temperature": 0.8}

  • Calm late-night narration

    Lowering exaggeration and guidance_scale gives a slower, flatter reading suited to a quiet narration line.

    {"text": "It's late, and the house is finally quiet. I could talk to you like this all night.", "exaggeration": 0.3, "guidance_scale": 0.3}

  • A spelled-out sigh instead of a tag

    Since bracketed tags are read aloud as words, a sound like a sigh is written out directly in the text rather than tagged.

    {"text": "Mmm... I didn't expect you to say that.", "exaggeration": 0.6}

Chatterbox API questions

Because there is one voice, the sliders are the only way to change how a line lands: keep exaggeration near the default for narration and push it toward 1 only for lines that are meant to be dramatic or breathless, since a whole script read at high exaggeration tends to sound uniformly intense rather than expressive. Write sounds out as words rather than tags -- a spelled-out 'mmm...' combined with a higher exaggeration value gets closer to a performed sound than any bracketed cue will. Because pricing is flat per request rather than per character, it is worth combining several short lines into fewer, longer requests where the script allows it.