Resemble Chatterbox

Resemble Chatterbox

Use Chatterbox when the production needs expressive AI speech, zero-shot/reference-voice behavior, local or on-prem deployment, or Resemble-hosted TTS with Chatterbox-family voices. Do not use it as a generic "celebrity impersonation" tool. Treat every reference voice as biometric/personal data and require documented rights before generation.

Volatile facts in this skill were verified on 2026-07-10 from official Resemble AI, GitHub, Hugging Face, and Resemble documentation sources listed at the end.

Choose the route first

Documented facts:

• Local/open route: Resemble publishes Chatterbox-family code and weights through the official resemble-ai/chatterbox GitHub repository and ResembleAI/chatterbox Hugging Face model card under MIT license. The local package installs with pip install chatterbox-tts and exposes Python classes such as ChatterboxTTS and ChatterboxMultilingualTTS.
• Hosted route: Resemble's synthesis API uses https://f.cluster.resemble.ai/synthesize for synchronous TTS and a streaming HTTP endpoint for chunked WAV. Requests authenticate with Authorization: Bearer YOUR_API_KEY.
• Hosted Chatterbox model versions documented by Resemble include Chatterbox, Chatterbox-Turbo, and Chatterbox Multilingual. Hosted Turbo is selected with the model value chatterbox-turbo; Resemble's model-version page also lists version codes tts-v4 and tts-v4-turbo.

Production decision:

• Prefer local/open Chatterbox when the project requires on-prem control, no vendor audio upload, custom batching, offline evaluation, or permissive model modification.
• Prefer hosted Resemble when the project needs managed voices, streaming API delivery, low ops burden, team voice libraries, billing support, or production scale without managing GPU inference.
• Prefer Turbo for low-latency agents or dialog loops where its documented paralinguistic tag support matters.
• Prefer Multilingual V3 for cross-language voice cloning or localization; use Single Language Pack models when a priority language or regional dialect needs tighter pronunciation quality.
• Prefer original English Chatterbox when the job is English narration/character voice with emotion exaggeration and local creative control, and multilingual coverage is not required.

Capability boundaries

Documented facts:

• Chatterbox is a family of open-source text-to-speech models by Resemble AI. The model card presents Chatterbox Multilingual V3 as a 500M-parameter general-purpose multilingual model, with Single Language Pack finetunes for selected languages and original Chatterbox for English creative controls.
• Official docs and model cards describe zero-shot/reference-voice synthesis. The open model card shows audio_prompt_path for conditioning on a different voice. Resemble marketing pages describe 5-second reference cloning for open Chatterbox/Turbo; hosted Resemble model-version docs list 10+ seconds as the dataset requirement for Chatterbox-family voice cloning.
• Chatterbox has emotion/intensity control through an exaggeration parameter in local examples. Resemble's hosted Chatterbox-Turbo docs describe native paralinguistic tags such as [cough], [laugh], and [chuckle]; the Turbo page also shows [sigh], [gasp], [cough], [laugh], [whisper], and [breath].
• Resemble's hosted synchronous endpoint accepts text or SSML, up to 3,000 characters, and can return wav or mp3; the hosted streaming HTTP endpoint accepts up to 2,000 characters and returns chunked WAV. The model-version page lists a 2,000-character limit for Chatterbox variants and notes some SSML tags are not supported, so treat the stricter endpoint/model limit as binding unless your current endpoint response proves otherwise.
• Public hosted docs say Chatterbox-Turbo is supported for Rapid English voices and Pre Built Library voices, and is not supported with Chatterbox Multilingual. Do not assume a custom multilingual voice can use model: "chatterbox-turbo" unless current account/API evidence proves support.
• Hosted API outputs include response metadata such as duration, synthesis duration, sample rate, and issues.
• Chatterbox outputs are described by Resemble as PerTh-watermarked by default. The open model card says generated audio includes PerTh watermarks that survive MP3 compression, audio editing, and common manipulations while maintaining nearly 100% detection accuracy; Resemble's PerTh materials describe imperceptible watermarking and later detection.

Boundaries and cautions:

• Do not promise perfect speaker identity, perfect accent transfer, or no hallucination. V3 is documented as reducing hallucinations, not eliminating them.
• Do not assume every official Resemble page has the same language list. The open GitHub/Hugging Face Chatterbox card lists 23 languages for general-purpose multilingual support; Resemble hosted/model pages have overlapping but not identical lists and newer marketing pages mention additional languages/Single Language Packs. Verify the target language against the exact model card or hosted model page at job time.
• Do not rely on PerTh watermarking as a substitute for consent, disclosure, or abuse review. Watermarking is provenance support, not permission.
• Do not remove or attempt to defeat PerTh. Resemble states that watermark removal is against intended use.
• Do not represent hosted API data as zero-retention unless the user has a Resemble contract or documentation explicitly covering that TTS route. Resemble's public privacy materials say personal data, including biometric/sensory data, is retained as needed to provide services; the Trust Center references retention and disposal controls based on customer requirements and contracts.

Reference voice handling

Documented facts:

• Local examples use audio_prompt_path to specify a reference recording.
• Resemble's Chatterbox pages describe zero-shot cloning from short reference clips with no fine-tuning.
• Hosted Resemble voice creation and terms require the user to have necessary rights, consents, and permissions; Resemble may require consent from the individual whose voice is cloned.

Production requirements:

1. Confirm voice rights before using any reference clip: owner identity, allowed use, geography/platform, duration, revocation terms, and whether the output must be labeled synthetic.
2. Keep reference clips clean: one consenting speaker, dry speech, no music, no heavy noise reduction artifacts, no overlapping voices, and enough phonetic/prosodic variety for the target use.
3. Match reference language and target language when possible. If transferring across languages, expect accent leakage and run accent/pronunciation QA.
4. For public, political, medical, financial, or celebrity-like voices, require explicit written authorization and escalate any ambiguity.
5. Store reference clips and generated outputs in access-controlled project storage. Avoid sending local-only reference assets to hosted APIs unless the user approves the vendor route.

Production heuristics:

• For local zero-shot: start with 5-20 seconds of clean speech if the exact local model card supports short prompts; if similarity or stability is weak, test longer clean clips before changing models.
• For hosted voice creation: plan for the documented 10+ seconds minimum and follow the account's current voice-creation workflow.
• Normalize input audio before inference: mono, speech-focused, no clipping, moderate loudness, and trim long silences. Preserve natural breaths if they define the speaker; remove unrelated room noise.

Prompting and delivery control

Documented facts:

• Local model-card tips say default settings around exaggeration=0.5 and cfg/cfg_weight=0.5 work for most prompts.
• Official tips say lowering cfg to about 0.3 can improve pacing for fast reference speakers; expressive/dramatic speech can use lower cfg and higher exaggeration such as around 0.7.
• Official tips warn that higher exaggeration tends to speed up speech, and lowering cfg can compensate with more deliberate pacing.
• Turbo supports text paralinguistic tags in hosted docs and Resemble materials.
• Resemble accepts SSML, but hosted model docs list unsupported Chatterbox SSML tags including <prosody>, <emotion>, <phonemes>, <substitutions>, <emphasis>, and <say-as>.

Production heuristics:

• Write TTS text as performance copy, not prose. Short clauses, natural punctuation, and intentional paragraph breaks matter more than dense instructions.
• Use capitalization and punctuation only where they serve the delivery; Resemble's Chatterbox page shows capitalization affecting emphasis, but overusing caps can make speech sound unnatural.
• Treat bracketed paralinguistic tags as performance events, not decorations. Use them sparingly, usually at sentence boundaries.
• Segment long scripts into semantically complete chunks. Keep each chunk under the strictest relevant limit, then crossfade or room-tone-match in post.
• For multilingual localization, localize the script idiomatically before synthesis. Do not feed literal translations and expect the model to fix rhythm.

Suggested starting points:

• Neutral narration: exaggeration=0.5, cfg_weight=0.5; clean reference; standard punctuation.
• Dramatic character read: exaggeration=0.65-0.85, cfg_weight=0.25-0.4; shorter sentences; test pacing.
• Fast reference speaker: reduce cfg_weight toward 0.3; add punctuation pauses before lowering speed in post.
• Cross-language clone: use Chatterbox Multilingual V3 or the relevant Single Language Pack; match language_id; QA accent and off-language phonemes.
• Agent/IVR Turbo: use hosted or local Turbo where available; keep utterances short; include only necessary tags such as [chuckle] or [breath].

Integration patterns

Local/open Python pattern

Use when the project can install Python dependencies and run inference locally. Pin package and model revisions for production.

Example:

import torchaudio as ta
from chatterbox.tts import ChatterboxTTS

device = "cuda"  # use "mps" or "cpu" only after performance testing
model = ChatterboxTTS.from_pretrained(device=device)

script = (
    "The first launch window opens at dawn. "
    "Keep your voice calm, precise, and just a little awed."
)

wav = model.generate(
    script,
    audio_prompt_path="approved_reference_voice.wav",
    exaggeration=0.55,
    cfg_weight=0.45,
)

ta.save("scene_01_narration.wav", wav, model.sr)

Why this works: it keeps reference-voice conditioning explicit, starts near documented defaults, and writes an uncompressed WAV suitable for editing and QA.

Local multilingual pattern

Example:

import torchaudio as ta
from chatterbox.mtl_tts import ChatterboxMultilingualTTS

model = ChatterboxMultilingualTTS.from_pretrained(device="cuda", t3_model="v3")

wav = model.generate(
    "Bonjour, votre colis est pret pour le retrait.",
    language_id="fr",
    audio_prompt_path="approved_french_reference.wav",
    exaggeration=0.5,
    cfg_weight=0.5,
)

ta.save("pickup_fr.wav", wav, model.sr)

Why this works: it selects V3 explicitly, supplies the language ID, and avoids cross-language accent transfer unless intentionally tested.

Hosted synchronous API pattern

Use when the user has a Resemble API key and a voice_uuid.

Example request:

curl -X POST "https://f.cluster.resemble.ai/synthesize" \
  -H "Authorization: Bearer $RESEMBLE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "voice_uuid": "YOUR_APPROVED_VOICE_UUID",
    "data": "Thanks for calling. I can help you check your appointment time.",
    "model": "chatterbox-turbo",
    "output_format": "wav",
    "precision": "PCM_16",
    "sample_rate": 48000,
    "use_hd": false
  }'

Why this works: it explicitly requests Turbo for lower latency/paralinguistic support and asks for edit-friendly WAV. Decode audio_content from base64, store duration/sample rate/issues, and retry only safe transient failures with backoff.

Hosted streaming pattern

Use streaming for interactive voice agents where time-to-first-sound matters. Keep data short, expect chunked WAV, and preflight whether the selected voice supports Turbo. Fall back to synchronous only if latency is acceptable and the user approves any experience change.

Production QA

Minimum review before delivery:

• Rights: reference voice, script content, and intended distribution are authorized.
• Model route: local/open vs hosted route is recorded with model variant, package/model revision or API model value, date, and settings.
• Text accuracy: no omitted words, hallucinated continuation, unexpected repetition, or off-script phrases.
• Voice quality: similarity is acceptable without impersonation risk beyond the approved use; no identity drift across chunks.
• Performance: pacing, emphasis, affect, breathing, and paralinguistic tags match the scene.
• Language: pronunciation, accent, names, numbers, and code-switching are correct for the audience.
• Audio engineering: no clipping, dropouts, long silences, robotic artifacts, room-tone mismatch, or discontinuities at chunk joins.
• Watermark/provenance: PerTh behavior is not disabled or stripped; provenance requirements are documented.
• Privacy/security: hosted inputs were approved for upload; local reference files are not leaked into logs or public artifacts.
• Accessibility and disclosure: where required by law, platform policy, or user brief, disclose that the voice is AI-generated.

Failure modes and repairs

• Off-prompt continuation or repetition: shorten the chunk, tighten punctuation, reduce expressive settings, try V3 for multilingual, and regenerate with a clean seed/run record.
• Too fast: lower cfg_weight, add punctuation breaks, avoid excessive exaggeration, and test a calmer reference clip.
• Too flat: raise exaggeration in small increments, rewrite text with clearer emotional beats, and choose a reference with the desired energy.
• Accent leakage in cross-language cloning: use a same-language reference, set the exact language ID, test cfg_weight changes, or move to a Single Language Pack.
• Paralinguistic tag spoken literally: confirm Turbo/tag support for the runtime, adjust tag syntax, or remove the tag and edit in a separate SFX/breath asset.
• API truncation or bad request: split text below the endpoint/model limit, remove unsupported SSML, and inspect issues returned by the API.
• Watermark concern after post-processing: run PerTh detection where available; keep a pre-master WAV and an exported distribution file for comparison.

Official sources verified on 2026-07-10

• Resemble AI Chatterbox GitHub repository
• ResembleAI/chatterbox Hugging Face model card
• Resemble Chatterbox model page
• Resemble Chatterbox Turbo model page
• Resemble Chatterbox Multilingual model page
• Resemble synchronous TTS API reference
• Resemble streaming TTS API reference
• Resemble authentication docs
• Resemble rate-limit docs
• Resemble model-version docs
• Resemble SSML reference
• Resemble pricing page
• Resemble Terms of Service
• Resemble Privacy Policy
• Resemble Trust Center
• Resemble PerTh GitHub repository
• Resemble PerTh model page