Text to speech (TTS) guide

Text to speech (TTS) guide

POST /v1/audio/speech is OpenAI-compatible: send JSON, get the audio file back as binary. Speech models include the ElevenLabs family (elevenlabs/eleven-v4, eleven-v4-turbo, eleven-v3, eleven-v3-conversational, eleven-multilingual-v2, eleven-flash-v2.5, eleven-flash-v2, eleven-turbo-v2.5, eleven-turbo-v2). The live list is GET /v1/models?output_modalities=speech. GET /v1/models returns text models only by default, so add ?output_modalities=speech (or all) to list them.

Parameters

The seven voice settings above are all optional; leaving them out changes nothing. A value out of range returns 400 and is neither sent nor billed. The response header X-Voice-Settings tells you whether your settings were used: applied = used; ignored = a fallback route answered, so your settings were not used (the audio is still produced); no header = you sent no voice settings. For example, stability 0 reads with more expression, stability 1 more evenly.

Voice rules

ElevenLabs models (elevenlabs/…) take one of these voice names (case-insensitive): george, sarah, adam, alice, bella, bill, brian, callum, charlie, chris, daniel, eric, harry, jessica, laura, liam, lily, matilda, river, roger, will. voice may be omitted for ElevenLabs models (default george). Any other value returns 400 invalid_voice. Your own custom voice (voice_…) is described in the Custom voices section below.

OpenAI-style voice names such as alloy, echo or nova only work with models that list them. Check the model page for each model's voices.

Custom voices (cloning) are being opened in stages: see the Custom voices section below (POST /v1/voices). Until they are enabled for your account, use one of the 21 preset names above or a public voice ID.

Built-in voices

Set voice to one of the names below (lowercase). Each voice has a short preview clip of our own, and GET /api/v1/audio/voices returns the full list (gender, age group, accent, style and preview URLs).

?lang= (en, zh-TW, zh-CN; default follows Accept-Language) only chooses the language of description. The preview URLs do not change with it: preview_urls.en (same as preview_url) is the English clip and preview_urls.zh (same as preview_url_zh) is the Chinese clip.

Emotion and action tags (audio tags)

Write tags in square brackets directly inside input, for example [laughs], [whispers], [excited], [sighs], [crying], [curious] or [softly]. They are sent to the model exactly as typed and shape how the following words are performed. Keep the tags in English even when the text is in another language.

curl https://api.bazaarlink.ai/v1/audio/speech \
  -H "Authorization: Bearer $BAZAARLINK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"elevenlabs/eleven-v3","input":"[laughs] That was actually funny. [whispers] Do not tell anyone.","voice":"george","response_format":"mp3"}' \
  --output audio-tags.mp3

Supported by elevenlabs/eleven-v3 and the elevenlabs/eleven-v4 family. Other speech models do not interpret tags and may read the brackets aloud, so try a short sentence first.

Tags count as input characters and are billed like the rest of the text.

Output formats

response_format defaults to mp3 (Content-Type audio/mpeg) when omitted; pcm (audio/pcm, raw 16-bit PCM) also works. The API accepts opus, aac, flac and wav, but ElevenLabs models currently reject them with 400. Save the file using the Content-Type response header.

Pricing

Speech is billed per character of input. The per-character price is on each model page under /models and in GET /v1/models as pricing.character (USD per character). Model pages show it per 1M characters.

Examples

curl https://api.bazaarlink.ai/v1/audio/speech \
  -H "Authorization: Bearer $BAZAARLINK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"elevenlabs/eleven-flash-v2.5","input":"Hello from BazaarLink.","voice":"george","response_format":"mp3"}' \
  --output speech.mp3

Common errors

Full parameter reference: Speech synthesis

Speech synthesis

Convert text to audio using an id from /v1/models?output_modalities=speech. Billing uses the model page price per million characters. The complete audio is buffered before returning binary data; incremental streaming is not supported. Routing control fields provider, transforms, route, service_tier, and preset are ignored.

curl https://api.bazaarlink.ai/v1/audio/speech \
  -H "$BAZAARLINK_AUTH_HEADER" \
  -H "Content-Type: application/json" \
  -d '{"model":"<speech-model-id>","input":"Hello","voice":"alloy","response_format":"mp3"}' \
  --output speech.mp3

# Voice settings (elevenlabs/* models; the X-Voice-Settings response header says applied or ignored)
curl -D - https://api.bazaarlink.ai/v1/audio/speech \
  -H "$BAZAARLINK_AUTH_HEADER" \
  -H "Content-Type: application/json" \
  -d '{"model":"elevenlabs/eleven-v4","input":"Hello","voice":"george","stability":0.4,"similarity_boost":0.8,"style":0.2,"use_speaker_boost":true,"language_code":"en","audio_effects":{"distance":0.3}}' \
  --output speech.mp3