Code-Generated Promo Videos (6): TTS Selection Guide — 31 Engines and Services Compared

Part 2 of this series used edge-tts to generate voiceovers, and Part 5 laid out the TTS technology evolution. This is the final installment — a complete selection guide for when your project outgrows edge-tts. What are the 31 TTS engines and services on the market, and how do you choose among them?

We organize the selection landscape into three layers: free open-source engines (run locally, zero licensing cost but with GPU deployment overhead), free cloud tiers (ready out of the box with usage limits), and paid services (ready out of the box + SLA + advanced capabilities). These are not strictly hierarchical — an open-source engine on self-hosted GPU can be cheaper long-term than paid cloud, while cloud services offer convenience that self-deployment can’t match. The choice ultimately depends on your budget, team capacity, and timeline.

Free TTS Comparison

Open-Source TTS Engines

All 12 engines below can be deployed locally at no licensing cost. They trade GPU compute for unlimited calling freedom — if your project has GPU access (even an RTX 3060 can run many of these), an open-source engine may be your most cost-effective long-term solution.

Note that “Commercial Use” refers to the model weights license, which may differ from the code license (e.g., F5-TTS uses MIT for code but CC-BY-NC for weights). Always verify final license terms before commercial deployment.

Key Engines in Detail

CosyVoice (Alibaba) — The top recommendation for local Chinese TTS. Apache-2.0 license fully permits commercial use. Supports Chinese, English, Japanese, Korean, and five Chinese dialects (Cantonese, Sichuan, Shanghai, Tianjin, Wuhan). Zero-shot cloning needs only 3 seconds of reference audio. First-packet latency as low as 150ms. Comes with WebUI, FastAPI, gRPC, Docker, vLLM deployment paths. Supports 14 fine-grained control tags ([laughter], [breath], etc.). 22k+ GitHub stars, active community.

F5-TTS (Shanghai Jiao Tong University) — Research-grade cloning quality. Flow Matching + DiT architecture, non-autoregressive, RTF ~0.15. Completely eliminates phoneme alignment requirements. Code is MIT but model weights are CC-BY-NC (not for commercial use, due to the Emilia training dataset). 14.5k+ GitHub stars, highly regarded in the open-source community.

ChatTTS — Purpose-built for dialogue scenarios. Trained on ~100K hours of Chinese and English data, optimized for conversational prosody including laughter and pauses. Minimum VRAM ~4GB, RTF 0.3 on RTX 4090. 39.3k+ GitHub stars, but the AGPL license prohibits commercial use.

Coqui XTTSv2 — Widest language coverage (17 languages). Zero-shot cloning from ~6 seconds of reference audio, 24kHz sampling rate. Note that Coqui Inc. shut down in January 2024; the community continues maintenance (~45k GitHub stars).

Piper — The go-to for edge devices. VITS architecture + ONNX export, CPU-first, runs on Raspberry Pi 4 with VRAM < 1GB. Supports dozens of pre-trained voices across many languages. Ideal for IoT and embedded scenarios.

Free Cloud Tiers

These services offer free usage quotas — ready to use without GPU management, but constrained by caps, rate limits, or time expiration.

Edge TTS is the most practical free option. Install with pip install edge-tts and it works immediately with no API key required. Based on Microsoft’s Neural TTS engine, it covers 100+ languages and 400+ voices with excellent quality. In Part 2 of this series, it served as the core voiceover tool — the gen_all_voices.py script, protected by 5 retries with exponential backoff, reliably produces hundreds of audio clips in batch.

But Edge TTS is not a silver bullet. It relies on an unofficial API that may encounter WebSocket 403 errors (Microsoft rate limiting), making it unsuitable for production high-concurrency scenarios. For production TTS integration, upgrade to the official Azure Speech Service SDK (paid + SLA-backed).

Paid TTS Comparison

When free options fall short — whether you need voice cloning, production SLA, broadcast-quality audio, or Chinese dialect coverage — it’s time for paid services. The paid TTS market splits into three pricing models: per-character cloud APIs (international cloud providers), monthly subscriptions (professional AI voice companies), and per-10K-character Chinese cloud services.

International Cloud Providers

Per-million-character pricing for mainstream international cloud TTS:

The $4/M chars standard tier has become an industry baseline — differentiation happens in premium quality tiers and ecosystem integration.

Amazon Polly offers 30+ languages and 90+ voices. Generative voices and Long-Form voices (asynchronous synthesis for up to 200K characters) are well suited for long-form audio content. Supports Speech Marks for lip-sync timestamps. New AWS users may receive up to $200 in free credits (since July 2025).

Azure Speech has the widest language coverage — 140+ languages/dialects, 500+ voices. Custom Neural Voice enables brand-specific voice training (minimum 300 recorded phrases). Emotional SSML extensions, Viseme output (facial animation sync), and Audio Content Creation Studio (browser-based visual tuning) make it a solid foundation for enterprise needs.

Google Cloud TTS offers the richest model selection: Neural2, Chirp 3 HD, Journey, Studio Voice, and more. The free tier resets monthly rather than expiring after a year — small-volume projects can use the standard tier indefinitely at no cost.

OpenAI TTS is known for its minimalist API design, integrating seamlessly with GPT-4o, Whisper, and the Assistants API. gpt-4o-mini-tts supports natural language instruction for voice style control, and the Realtime API enables end-to-end sub-second voice conversations. At $15/M chars, it costs roughly 1/10 of ElevenLabs, but voice selection (6+) and quality options are limited.

Professional AI Voice Companies

Monthly subscription services targeting content creators and professional voice cloning needs:

ElevenLabs is the undisputed voice quality leader, with an MOS of ~4.3 — broadcast-grade quality. 70+ languages, 11,000+ voice library, Flash v2.5 latency ~75ms. Voice cloning is its crown jewel: instant cloning (30 seconds of audio) on Starter+, professional cloning (30 minutes of training audio) on Creator+. Dubbing Studio supports 70+ languages with lip-sync, and ElevenAgents provides STT+LLM+TTS+telephony in one package. Sound Effects and Music generation are also available.

Play.ht boasts the largest voice library (800+), with one free voice clone included even in the free tier. Low-latency real-time API (10 req/sec rate limit), suitable for content creation that requires frequent voice switching.

Murf.ai positions itself as an all-in-one content creation platform with 200+ voices, deep Canva and Google Slides integration, built-in timeline editor, video sync, and Studio Sound enhancement.

WellSaid Labs emphasizes ethical AI use and legal compliance with a thorough voice licensing framework, suitable for regulated industries. Custom Lexicons support domain-specific pronunciation.

Descript Overdub stands apart — it embeds TTS directly into a text-based audio/video editor. Pro tier supports unlimited Overdub, and features like Studio Sound and AI Green Screen make “editing video like editing text” a reality.

China Domestic Services

Per-10K-character pricing for Chinese cloud TTS providers:

The Chinese TTS market is dominated by cloud vendors with fierce price competition — the basic tier has essentially settled at ¥0.3-0.4 per 10K chars. Differentiation happens in large-model tiers, voice cloning, dialect coverage, and free quota.

Alibaba Cloud builds on the Sambert model and the open-source CosyVoice LLM, with a focus on e-commerce scenarios (customer service voices). New users get 1M chars/month free for 3 months. Best for teams already on Alibaba Cloud ECS.

Tencent Cloud offers four interface tiers: basic/premium/LLM/ultra-real LLM, with tiered pricing (lower unit cost at higher volume). The ultra-real LLM at ¥4.9-6.5/10K chars delivers the highest quality. Basic voice free quota is the largest at 8M chars/month. Voice cloning supports one-phrase quick cloning.

Baidu AI Cloud supports dialects (Sichuan, Cantonese) and emotional voices (happy, sad). Premium quality outputs 24kHz audio. Free quota of 2M chars/month for 1 year makes it a solid pick for cost-sensitive projects.

iFlytek has the most comprehensive dialect coverage — 16 dialects including Henan, Northeast, and more. Its offline SDK is mature and stable, deployable to 8051 MCUs and ESP32 embedded devices. Permanent free quota of 100K chars/month (600K after verification). Spark TTS large model is being rolled out.

Volcano Engine (ByteDance / Doubao) offers the Doubao TTS model 2.0 based on Seed-TTS, plus an end-to-end ultra-lifelike real-time voice model. Voice cloning 2.0 and fine-grained emotion control are key differentiators. Priced at ¥5/10K chars, it’s best for teams within the ByteDance ecosystem (Douyin, Feishu).

Core Capability Matrix

This matrix covers 18 major providers across key dimensions — your fastest reference for cross-vendor comparison:

Application Scenarios

TTS is evolving from a tooling capability into the underlying infrastructure for voice interaction. Here are the 8 most active application areas:

Real-time voice dialogue (GPT-4o Voice, Doubao Realtime Voice, Sesame CSM) is the hottest direction for 2024-2025. Unlike traditional STT+LLM+TTS pipelines, next-generation systems reason directly in the speech modality for sub-second response with emotional adaptation.

Film dubbing and content localization (ElevenLabs Dubbing Studio) automatically translates and voices content into 70+ languages with lip-sync while preserving original emotion and intonation, dramatically reducing traditional dubbing costs.

Audio content production (audiobooks, podcasts, news narration) is now highly automated — multi-character voice switching, emotional control, and Studio Sound enhancement are all standard features.

Accessibility (Speechify, Apple VoiceOver) reads screen content for visually impaired users across multiple platforms; improved naturalness significantly enhances user experience.

Smart customer service and IVR deeply integrates emotion recognition and affective response, with Custom Neural Voice for branded voices and Viseme output for virtual human facial animation.

Gaming and virtual humans use dynamic speech generation and Viseme facial sync to drive NPCs and digital avatars.

Education and language learning leverage TTS for text reading, pronunciation demonstration, personalized voice tutors, and spoken language assessment.

Embedded and IoT (iFlytek offline SDK, Piper) run TTS offline on smart speakers, vehicle systems, and Raspberry Pis, ensuring privacy and low power consumption.

Selection Guide

Decision Tree

Start by dichotomizing your needs: personal/prototype development should prioritize free options to minimize upfront cost; production/commercial environments should choose paid services for SLA guarantees and advanced capabilities.

flowchart TD
    A["TTS Selection"]:::start
    A --> B["Free / Personal"]:::process
    A --> C["Paid / Production"]:::special
    B --> D["Edge TTS<br/>Quick & Free"]:::output
    B --> E["CosyVoice<br/>Local Deploy OK"]:::output
    C --> F["ElevenLabs<br/>SLA Quality"]:::output

    classDef start fill:#FF9800,color:#fff
    classDef process fill:#2196F3,color:#fff
    classDef special fill:#9C27B0,color:#fff
    classDef output fill:#4CAF50,color:#fff

Decision tree walkthrough: The root is “TTS Selection” — your first split is budget. The free path has two exits: Edge TTS for zero-cost rapid prototyping (ideal for proof-of-concept and personal projects), or open-source engines like CosyVoice for local deployment (when you need voice cloning or Apache-2.0 commercial licensing). The paid path leads to ElevenLabs (broadcast-grade quality) or cloud provider services (SLA-backed, security-compliant). These paths aren’t mutually exclusive — a common migration path is: validate with Edge TTS during prototyping, then switch to CosyVoice self-deployment or ElevenLabs for production.

Selection by Use Case

Selection by Technical Need

Pricing Model Reference

The pricing model of a paid TTS service directly impacts your selection. Per-character billing (AWS Polly, Azure, Google Cloud, OpenAI, Chinese clouds) suits on-demand scenarios with unpredictable volume — no fixed cost but fully elastic. Monthly subscription (ElevenLabs, Play.ht, Murf.ai, WellSaid) fits content creation teams with stable monthly output — the subscription cap protects against unexpected spikes. Per-minute billing (ElevenLabs ElevenAgents at $0.08/min) targets real-time dialogue and telephony, aligning billing granularity with usage. Tiered subscription (IBM Watson Lite/Plus/Premium) segments features by company size. There is also iFlytek’s permanent free quota (100K chars/month) — a cost-conscious option for small-scale long-term use that works indefinitely without any payment, albeit with feature and service-level limitations.

⚠️ Prices and figures are current as of the report date (2026-07). Consult official websites for the latest pricing.

Series: ① Overview & Remotion ② edge-tts Voiceover ③ numpy BGM ④ ffmpeg Muxing ⑤ TTS Landscape ⑥ TTS Selection Guide