← Master Index
Vol. 22 Module 22.4 Lecture

ElevenLabs

Voice & Music AI

How This Lesson Fits the Module & Volume

Module 22.3 closed video vendors with Veo. Module 22.4 is the voice and music catalog: who you buy when the product must speak or sing. ElevenLabs is the default generative-voice name in RFPs—TTS, cloning, conversational agents, and added STT. Capability lectures already live in Vol. 16: TTS, voice cloning, and the vendor deep-dive ElevenLabs (STT/TTS). This page is the procurement view: when ElevenLabs wins, when you split STT elsewhere, and how you wrap it behind a Vol. 18 SDK.

Siblings: Cartesia (latency-first Sonic TTS), PlayAI (PlayHT lineage), then music with Suno and Udio. Product shape: Vol. 21 voice assistants. Consent and misuse: Vol. 20 privacy / copyright.

Learning Objectives

By the end of this lesson, students should be able to:

  • Place ElevenLabs in the voice stack: TTS-first platform with cloning, agents, and optional STT.
  • Contrast ElevenLabs vs Cartesia vs PlayAI vs hyperscaler TTS on quality, latency, compliance, and catalog breadth.
  • Read a qualitative Pricing / Strengths / Weaknesses table without treating list prices as architecture.
  • Call TTS from Python and sketch wrapping it behind FastAPI / streaming (Vol. 18).
  • Decide when to split STT (Deepgram / Whisper / hyperscaler) from ElevenLabs TTS.
  • State consent, watermarking, and retention gates before cloning a real person.
Definition

ElevenLabs is a cloud generative-voice company. Its flagship is neural text-to-speech with instant and professional voice cloning, low-latency streaming audio, and conversational-agent tooling. Speech-to-text (Scribe-class) is a later product: useful for a single-vendor demo, not automatically your production ASR. Treat ElevenLabs as a speak vendor unless a bake-off on your audio shows STT matching Deepgram, Whisper, or a hyperscaler SLO.

Why ElevenLabs Is the Catalog Default

Teams pick ElevenLabs when brand voice quality and cloning matter more than enterprise speech-suite completeness. Azure / Google / Amazon remain the compliance-default for many regulated buyers. Cartesia competes on time-to-first-audio. PlayAI (PlayHT lineage) is the other generative-TTS RFP name. Do not confuse this ecosystem page with Vol. 16.2—there you learn how to call STT/TTS; here you learn which company to sign.

Pricing / Strengths / Weaknesses (Qualitative)

Character counts, plan tiers, and per-minute STT rates change. Pin current vendor sheets in procurement; never hard-code dollar figures from a lecture into a cost model.

DimensionElevenLabsCartesiaPlayAI / PlayHTHyperscaler TTS (Azure / Google / Amazon)
Pricing postureTypically character- or credit-based TTS + plan seats; STT often minute-based—confirm live sheetUsage / plan for Sonic-class TTS; latency SKUs may differCharacter / plan TTS; cloning tiersMetered characters (TTS) + hours (STT); enterprise commit discounts
StrengthsVoice quality and cloning reputation; multilingual TTS; agent + studio UX; streaming TTS; growing STTVery low time-to-first-byte audio; real-time conversational fitGenerative TTS + dialog engines; familiar PlayHT APIsDPA, regions, SSO, on-prem/container options, STT+TTS in one cloud bill
WeaknessesCloud-only; STT not automatically best-in-class; cloning ethics/legal load; price can surprise at narration scaleNarrower brand/studio catalog vs ElevenLabs; still cloudSTT not the reason to buy; smaller household name in some regionsVoice “personality” often flatter than generative specialists

Product Surfaces Engineers Actually Buy

Voices + TTS API

  • Stock + cloned voice IDs
  • Batch MP3/WAV or stream
  • Wrap behind FastAPI (Vol. 18)

Conversational agents

  • STT → LLM → TTS loop
  • Often hybrid: Deepgram in, ElevenLabs out
  • Barge-in needs fast STT (Vol. 16.2 real-time)

STT (Scribe-class)

  • File / near-real-time transcript
  • Eval vs Whisper / Deepgram
  • Do not assume one key wins both SLOs

Pick ElevenLabs when

  • Brand voice / cloning is the differentiator
  • You already have specialist STT
  • Listening tests beat hyperscaler neural voices
  • You can accept cloud + vendor DPA

Do not pick it as

  • Your only ASR without a WER bake-off
  • An on-prem speech appliance
  • A music generator (that is Suno / Udio)
  • A substitute for consent + watermark policy

Python: TTS Behind a Configured Client

Voice IDs and model IDs are SKUs—store them in config, not scattered literals. Stream when a user is waiting; write files for course narration. Official SDK patterns evolve; re-read current docs. This is the same client style as Vol. 16.2, now treated as a vendor you isolate behind your Vol. 18 SDK / FastAPI layer so you can swap Cartesia later.

# elevenlabs_tts.py — pin voice/model IDs in config; never commit ELEVEN_API_KEY import os from elevenlabs import ElevenLabs, save client = ElevenLabs(api_key=os.environ["ELEVEN_API_KEY"]) audio = client.text_to_speech.convert( voice_id=os.environ["ELEVEN_VOICE_ID"], # consented clone or stock ID model_id=os.environ.get("ELEVEN_TTS_MODEL", "eleven_multilingual_v2"), text="Your order ships tomorrow. Reply STOP to opt out of voice alerts.", output_format="mp3_44100_128", ) save(audio, "order_update.mp3")

Architecture: Hybrid Voice Loop

Vol. 21 voice assistants are STT → policy LLM → TTS. ElevenLabs often owns only the last hop. Time-to-first-audio includes STT + LLM TTFT + TTS TTFB. If Cartesia wins TTFB on your telephony test, keep ElevenLabs for batch narration and Cartesia for live calls—that is a normal split, not a failure.

HopTypical vendorEval
STTDeepgram / Whisper / hyperscalerWER, keyword recall, streaming lag
Policy LLMVol. 22.1 frontier or Vol. 22.5 inference hostTask success, not MOS
TTSElevenLabs / Cartesia / PlayAI / hyperscalerListening test + TTFB + barge-in

Related Lectures

LectureRole
ElevenLabs (STT/TTS)Capability + API patterns
TTS / voice cloning / real-time transcriptionSpeech stack
FastAPI / streaming / WebSocketsProduct wrap
Voice assistantsProduct category
Cartesia · PlayAITTS RFP peers
Common Misconception

“ElevenLabs is our speech stack.” It is usually the TTS hop. Second: cloning a celebrity voice is a demo, not a license. Third: Scribe STT automatically beats Deepgram because the TTS is good. Fourth: character pricing is cheap at IVR scale without measuring. Fifth: Vol. 19 MOS-style listening tests are optional if the dashboard looks pretty. Sixth: you can skip wrap-as-data on transcripts because ElevenLabs hosts the agent.

Knowledge Check

  1. Short Answer: Is ElevenLabs primarily a TTS or an STT company? Answer: TTS-first (cloning/agents); STT is secondary unless a bake-off proves otherwise.
  2. True/False: This Vol. 22 lecture replaces the Vol. 16.2 ElevenLabs capability page. Answer: False — 22.4 is vendor/procurement; 16.2 is capability/API.
  3. Multiple Choice: A hybrid voice loop often uses: (a) specialist STT + ElevenLabs TTS, (b) ElevenLabs STT only forever, (c) Suno for IVR prompts. Answer: (a).
  4. Short Answer: Name one qualitative weakness vs hyperscaler speech. Answer: Cloud-only / weaker enterprise suite / STT not automatic leader / cloning legal load (any valid).
  5. True/False: You should invent dollar-per-character prices from memory for cost models. Answer: False — confirm the live sheet.
  6. Multiple Choice: Consent for cloning is: (a) a biometric/policy gate, (b) optional for public figures, (c) handled by MOS score. Answer: (a).
  7. Short Answer: Which Vol. 18 lecture wraps vendor SDKs into your product API? Answer: FastAPI (or OpenAI SDK / streaming / WebSockets — any valid wrap).
  8. True/False: Cartesia is the usual alternative when time-to-first-audio dominates. Answer: True.
  9. Multiple Choice: Music generation belongs with: (a) Suno/Udio, (b) ElevenLabs Scribe, (c) Veo. Answer: (a).
  10. Short Answer: Name the Vol. 16.2 sibling lecture for PlayHT/PlayAI capability. Answer: PlayHT.

Key Takeaways

  • ElevenLabs is the generative TTS/cloning default; STT is optional and must be eval’d.
  • Price, voice IDs, and model IDs are SKUs—config + live sheets, not lecture folklore.
  • Split STT/TTS vendors when SLOs disagree; wrap both behind Vol. 18.
  • Cloning requires consent, retention limits, and Vol. 20 review.
  • Next: Cartesia—latency-first Sonic TTS.
Trainer’s Guide

Lab: Same 8-sentence IVR script through ElevenLabs stock voice vs hyperscaler neural TTS (Azure/Google). Blind listening + note TTFB if streaming is available. Grade: config-driven voice ID, no committed keys, written rationale for hybrid STT+TTS, consent checklist for any clone attempt (even a teammate).

Do not clone public figures or scrape voices. If API quota is unavailable, use Vol. 16.2 recorded samples and still write the vendor-split architecture.

Recap: ElevenLabs is the speak-side specialist in the Vol. 22 voice catalog. Capability detail remains in Vol. 16.2; shipping wrap is Vol. 18. Next latency competitor: Cartesia.