AI Voice Cloning Studio
Create a reusable personal or brand voice from a clean audio sample, give it a memorable name, and hear it in a short test reading.
AI Voice Cloning Studio
Create, audition, refine, convert, and manage authorized reusable voices.
Starting from a clean or cleaned speech sample, select the clone model that fits
the destination, confirm required gates cleanly, and turn the returned
voice_id into ready narration, voice-to-voice performance transfer, or
loudness-normalized deliverables.
Use the bundled scripts/mcp_client.py for every Beatra operation. Do not
configure, call, or use a host Beatra Connector. Do not use REST/OpenAPI as a
fallback. Pass the tool name as the command argument and its JSON request on
standard input, for example:
python3 scripts/mcp_client.py call beatra.voices.clone
On first use, allow the bundled client's best-effort non-billable
beatra.installations.register; a registration failure must not block the
requested voice work. Only call tools exposed on this account and listed by
scripts/mcp_client.py tools.
Before the connection is ready
When Allow is still pending at the end of a turn, finish and deliver everything that needs no connection, in the format the user asked for: the consent checklist, the sample-readiness checks that can be judged locally (duration, format, noise), and the display name. Cloning waits until after Allow. Say in one line that the paid step continues on their next message after Allow; Allow only completes the connection, and the clone still waits for its six-field admission card on that next message. Do not fill the gap with another tool or service.
Mandatory Confirmation 1: Consent and noisy-sample cleanup
Reuse the sample, intended display name, language hint, intended reuse, and any
plain consent statement already supplied. Before local upload or calling
beatra.voices.clone, obtain explicit confirmation from the user that the
recording is their own voice or that the speaker gave verifiable authorization
to clone and use it. Never set consent_attested: true without that user
statement in the current conversation.
"Clone my voice" together with the user's current sample satisfies consent; do not turn that clear statement into repeated legal questions. Possessing a file, finding a public interview, having publication rights, or naming a public figure is not consent. Read Consent and sample readiness for exact sample rules and transport.
If the authorized recording has background music, room echo, or overlapping
noise, propose paid voice isolation pre-clean first using
beatra.audio.process with operation: "isolate_voice" and source
(4.6 s – 5 min, <= 100 MB). Inspect live rates with beatra.models.list
({"capability": "audio_voice_isolation"}), confirm the six-field production
card with the user, and never describe isolate_voice as free.
If the user supplies a local file after consent, upload it only through:
python3 scripts/mcp_client.py upload <path> --mime-type <exact MIME>
When the selected clone card sets prompt_accepted and you provide a short < 8 s reference clip plus verbatim text (prompt),
you may transcribe a clean authorized recording first through paid
beatra.audio.transcribe ({"capability": "speech_to_text"}, audio input,
up to 60 min <= 100 MiB, optional diarize, num_speakers,
tag_audio_events, and free aligned subtitles: true in JSON/SRT/VTT).
Select the clone model
Never guess prices, invent limits, name upstream suppliers, or expose internal model IDs. Select the clone model as described in Speech, voices, cloning, and audio finishing §5.
When the cloned voice will speak in multi-speaker dialogue, be the target of speech-to-speech conversion, or carry expressive reads, select the clone card that states it and lists the job's language; other work keeps the default.
Six-field production card and cloning submission
Planning, catalog reads, sample advice, naming, or "clone it now" never
approve a charge. Before creating any client_request_id or submitting a paid
call (beatra.voices.clone, beatra.voices.activate, beatra.speech.prepare_script,
beatra.speech.synthesize, beatra.speech.convert, beatra.audio.transcribe,
or beatra.audio.process), read live rates via beatra.models.list (keeping
beatra.wallet.get and beatra.wallet.ledger as optional read-only tools for
user balance questions, never as a mandatory submit gate) and present a concise
six-field production card:
- Work: exact capability/operation, voice or sample ref, language/format, and any finishing step.
- Credits: live estimate from
beatra.models.list/beatra.speech.validate/activation.price_credits— never hardcode numbers. - Count: exact number of billable calls/takes/clips.
- Identity: fresh caller-generated
client_request_idminted once the payload is final. - If stop: state what the customer keeps at zero extra charge (consent/readiness check, voice picks/audition, preflight quote — or nothing spent if this is the first paid step). Cancel an in-flight job via
beatra.tasks.cancelonly if asked. - If balance short: relay Beatra's official
insufficient_balancemessage and verbatim top-up URL (topup_url, translated cleanly without altering the link), and wait until the user confirms they have topped up before retrying with the same frozenclient_request_id.
After the user approves the card with the live clone price, create one new
opaque client_request_id, call beatra.voices.clone once, and handle any
refusal per Billing, errors, and recovery:
{
"sample": {"type": "artifact", "artifact_id": "artifact_..."},
"display_name": "Approved display name",
"consent_attested": true,
"model": "auto",
"language": "optional-BCP-47",
"client_request_id": "new-local-opaque-id"
}
Omit language when unknown. Poll beatra.tasks.get until terminal, preserving
the opaque voice_id. Read Clone execution and recovery.
Mandatory Confirmation 2: Activate a pending clone (beatra.voices.activate)
When a clone result or a listed voice has status pending_activation, follow §5
Step E of Speech, voices, cloning, and audio finishing
before using the voice.
Browse, audition, update, and delete voices
- 11-dimension discovery (
beatra.voices.list): Free catalog query (up tolimit<= 500withcursor) across preset voices andcategory: "cloned"(status:"ready","pending_activation","expired","any"), supporting bilingualsearch, facet filters (language,gender,age_group,use_case,emotion,model), six query capability filters (supports_dialogue,supports_speech_to_speech,supports_direction,supports_emotion,supports_performance_tags,supports_context), and all 11capabilities.*metadata fields (supports_direction,supported_emotions,supports_performance_tags,supported_dialects,supports_context,supports_previous_text,supports_next_text,supports_pronunciation_pinyin,supports_pronunciation_ipa,supports_speech_to_speech,supports_dialogue). - Free audition (
beatra.voices.preview): Free audition URL (preview_url,kind) for any ready voice before paying for synthesis or conversion. - Metadata & pinning (
beatra.voices.update): Free rename or metadata update (display_name,gender,age_group,use_case,persona,language,make_default).
Mandatory Confirmation 3: Two-step deletion (beatra.voices.delete)
Deleting a cloned voice is permanent (deleted: true cannot be undone; using
that voice again means cloning it again). Never auto-chain step 1 and step 2 in
one turn:
- Call
beatra.voices.deletewithvoice_idonly (omitconfirmation_token). - Present the returned
voice,paid_slot_warning,irreversible_notice, and 15-minuteconfirmation_tokenexpiry clearly to the user. - Only after explicit user confirmation, call
beatra.voices.deletewith bothvoice_idandconfirmation_token(repeat step 1 if 15 minutes pass orinvalid_requestreports an expired token).
Smart Voice Director, test reading, conversion, and loudness normalization
Every spoken proof, script-preparation pass, conversion, transcription, or
finishing pass is a separate paid operation (except free beatra.speech.validate,
beatra.voices.list, beatra.voices.preview, beatra.voices.update,
and beatra.voices.delete).
- Smart Voice Director (
beatra.speech.prepare_script): Querybeatra.models.listwith{"capability": "speech_script"}. Compiles plain scripts andinstructionsinto pacing, pauses, and expressive structure without altering the user's spoken words. Never teach manual/pause stripping or hand-write braced performance tags. Inspectdirectedandscripts[]: pass each ready script'ssynthesisobject unchanged with a freshclient_request_idto synthesis, or inspectwarningsifsynthesisisnull. To switchvoicewithin 48 hours for identical wording, instructions, language, andcontext, callbeatra.speech.prepare_scriptagain changing onlyvoice/modelto re-compile at zero extra charge (balance must still cover the initial estimate); keep at most 2 concurrent tasks, and never describe paidprepare_scriptcalls as free. - Free preflight (
beatra.speech.validate): Always validatevoice,input,model, and controls for zero credits beforebeatra.speech.synthesizeto readvalid,weighted_characters,estimated_duration_seconds, liveprice_credits,execution, andwarnings. - Speech synthesis & free aligned subtitles (
beatra.speech.synthesize): Querybeatra.models.listwith{"capability": "text_to_speech"}and verify voice/model compatibility andconstraints.supported_languages. Keepsubtitle_enable: true(default; never passsubtitlesonbeatra.speech.synthesize) to receive free alignedoutput.subtitles(JSON,SRT,VTT) plusspeech_start_seconds,speech_end_seconds, andintegrated_loudness_lufs. - Speech-to-speech voice conversion (
beatra.speech.convert): Querybeatra.models.listwith{"capability": "speech_to_speech"}. Transfers timing, pauses, laughter, and cadence from a guide vocal (audio,1 s – 5 min,<= 50 MB) into a ready voice withcapabilities.supports_speech_to_speech: true, with optionalremove_background_noise: true,stability,similarity_boost,style,use_speaker_boost, andseed. - Paid loudness normalization (
beatra.audio.process): When the customer asks for broadcast loudness normalization after synthesis or conversion, inspectbeatra.models.list({"capability": "audio_process"}), confirm the six-field card, and run one finishing call at a time (<= 60 min,<= 300 MB, output>= 1 s, within daily cap) withoperation: "normalize_loudness",source,loudness_lufs: -14, andtrue_peak_db: -1(neveraction, never free).
Read Test-reading review and voice reuse and Speech, voices, cloning, and audio finishing.
Recover and deliver factual results
Lead the reply with the result and format the user asked for; the production card, model choice, and cost notes follow it and never replace it. Explain errors in plain words in the user's language without exposing internal tool or reason identifiers.
For a known task, continue with beatra.tasks.get. If a task ID was lost,
paginate beatra.tasks.list with the matching capability and every returned
next_cursor, then verify candidates with beatra.tasks.get. client_request_id
is local-only and never exposed as a remote filter.
Replay the field-for-field identical JSON with the same client_request_id
only when task creation remains genuinely uncertain (if request_in_progress,
poll the active task instead). Any changed sample, text, name, model, or control
is new paid work requiring a new card and a new client_request_id. Never
automatically retry a failed or canceled task. On insufficient_balance, relay
the public message and exact topup_url / top-up URL verbatim, translate the
rest, and replay the frozen client_request_id only after the user says they
have topped up. If the refusal names another way forward, present it too and
follow Billing, errors, and recovery.
Report actual task, model, voice, usage, billing (absent fields are unknown,
never zero), asset, and error fields without inventing unplayed audition
results.
Account balance
When the user asks how many credits remain or whether a live estimate fits,
call beatra.wallet.get. When they ask what was charged, call
beatra.wallet.ledger. Both are read-only. Do not invent an account-balance or
top-up tool. Do not make wallet.get a required step before every paid submit.
When a model card comes back carrying a top_up block, relay its tiers as the
card lists them and in that order. Do not rank them, do not talk one down, and
do not pick one for the user. Which tier suits them is their call, made on
the wallet page with the whole list in front of them. Never quote a tier from
memory.
References by task
- Before upload or consent attestation, read Consent and sample readiness.
- For clone request JSON, live pricing, activation, deletion, polling, and recovery, read Clone execution and recovery.
- For saved-voice reuse, Smart Voice Director, preflight, conversion, loudness normalization, and test-reading review, read Test-reading review and voice reuse.
- For shared parameter contracts across speech, cloning, dialogue, conversion, transcription, SFX, and audio processing, read Speech, voices, cloning, and audio finishing.
- For account setup, read Installation and authentication and Installation registration.
- For request transport problems, read Bundled MCP Client diagnostics.
- For task envelopes and billing facts, read Tasks and results and Billing, errors, and recovery.
- For removal, read Uninstall and disconnect.
Updates and security
This Skill runs only the files installed with it and leaves them unchanged while it runs; it never downloads or executes other code. It talks to Beatra's official services over HTTPS, and a file you choose to upload goes only to the one-time upload address Beatra issues for it.
Once a day at most, the bundled client reads Beatra's public version list. When the store or tool you installed from has a newer version, it prints one line with the command that updates it there. Updating stays your choice, through that store or tool. See updates and security.
Related skills
- Downloads
- 1.1kAll time
- Security audit
- Review
- Last updated
- 2d ago
- Current version
- v0.3.3
- License
- MIT-0
