Text-to-Speech Tools Compared: A Buyer’s Guide

Start with the job, not a total ranking

Text-to-speech products can sound similar in a short demo while solving very different jobs. A reading app must ingest documents and keep a listener's place. A narration workspace needs script revision, project organization, export control, and commercial rights. A live application needs streaming, predictable concurrency, and failure handling. A cloning workflow adds consent, identity, and disclosure requirements before voice quality even matters.

ElevenLabs is the broad default starting route when the job is still mixed or undecided. Its official product surfaces connect browser generation, long-form Studio projects, voice cloning, and API use through one workspace and shared credits. That breadth makes it a practical first trial, not a universal winner. Start elsewhere when the job is already specific: Speechify for reading, Murf AI for editor-led business narration, Cartesia for real-time API speech, Fish Audio for cloning with a separately metered API, and Unreal Speech for character-heavy API workloads.

Decision table

Reading, creator, and business narration

Speechify Reader is the clearest route when the desired output is listening rather than a media asset. It accepts documents and webpages, supports scanning printed text, and follows the user across supported devices. Its free route is suitable for testing the reading habit; Premium is the upgrade when voice quality, playback speed, OCR, integrations, or higher-use convenience matters. The important boundary is product identity: Reader, Studio, and the developer API have separate pricing and should not be treated as successive tiers of one subscription.

Move from Speechify Reader to Speechify Studio only when the job changes from consuming text to producing narration. Studio meters creation in credits per second and adds editing, cloning, and commercial-use paths. Embedded or automated speech belongs to the separate API, which meters TTS by billable characters. A buyer who only wants articles read aloud should not pay for a creator workflow, while a publisher should not assume a Reader subscription grants Studio export or API rights.

ElevenLabs is the better first trial for a creator who expects the workflow to expand. Studio can import text, documents, or URLs, arrange narration on a timeline, and export projects or chapters as audio. The same workspace can also supply API keys and cloning tools, so a prototype does not immediately force a new vendor or a separately branded product.

Free access is for auditioning; the paid boundary begins when commercial use, instant cloning, or production volume matters. Professional Voice Cloning is a further upgrade and is restricted to a verified clone of the user's own voice.

Murf AI fits a different production pattern: a human editor works through script blocks, pronunciation, pauses, emphasis, media, music, and a timeline before exporting a reviewed deliverable. Voice-only, mixed voice-and-music, video, and subtitle outputs make it useful for training, sales, internal communications, and other business narration where the file and review process matter as much as raw synthesis. Paid Studio access is the practical starting point because free workspaces cannot download.

Do not mistake a higher Murf self-serve tier for a collaboration plan. Official workspace guidance places multiple editors, shared folders and permissions, comments, and simultaneous collaboration in Enterprise. Automation is another boundary: Murf's API uses a separate key and character pricing rather than consuming a Studio project's production allowance. Exact Studio project and hour limits should be checked in the live checkout because Murf's pricing and help surfaces do not agree on every quota.

Real-time API, cloning value, and low-cost volume

Cartesia is the strongest specialist route here for real-time API TTS. Its Bytes, server-sent events, and WebSocket endpoints all stream audio, but they fit different interaction patterns. WebSockets accept partial text, reuse a connection across turns, return timestamps, and multiplex contexts, which is the useful shape for an agent or other interactive application. Cartesia describes Sonic 3.5 as sub-90ms, but that is a vendor model claim without a published percentile or network condition. Measure complete time to first playable audio from the intended deployment region.

Capacity is unusually explicit. The self-serve plans publish 2, 3, 5, and 15 concurrent TTS contexts, and an HTTP request or unique WebSocket context counts toward the limit. That makes the upgrade decision testable: load the real turn pattern, observe peak active contexts and 429 responses, then buy the smallest tier with headroom. Commercial use begins on a paid route, while organizations, professional voice cloning, SSO, and negotiated compliance introduce separate upgrade gates.

Fish Audio is the more interesting route when cloning flexibility and API unit economics sit in the same decision. Its cloning documentation supports both persistent reusable voices and inline reference audio for a single request. The API has no monthly subscription minimum and prices supported TTS models at $15 per million UTF-8 bytes, while published concurrency rises from 5 to 15 to 50 as prepaid spend crosses stated thresholds. That is straightforward for developers, but byte billing means the same visible character count can cost differently across scripts and encodings.

Fish's browser plans must be budgeted separately from its API. They use subscription credits and estimated audio minutes, not API bytes. The vendor's pricing card and terms also disagree about free-tier commercial use, so the safe publication boundary is a paid plan plus a rights check unless Fish confirms otherwise in writing. Choose Fish when reusable or instant cloning and direct API access are central; choose a simpler non-cloning API when identity management would add risk without product value.

Unreal Speech is the low-cost API route for predictable character volume. Its pricing page includes 250,000 free characters and advertises a Basic introductory price of $4.99 per month for the first six months with 3 million characters; higher tiers publish both larger packages and lower overage rates. The API separates short raw-audio streaming, medium timestamped synthesis, long asynchronous jobs, and WebSocket audio with word timing. That endpoint range is useful for bulk narration, accessibility audio, and other workloads where cost and throughput matter more than a cloning ecosystem.

Treat the introductory Unreal price as temporary, calculate the post-promotion and overage case, and verify capacity in the dashboard or with sales. The public request-rate table belongs to an older API version and does not map cleanly to the live plan lineup. Unreal's product page also says voice cloning is not available, which makes it an anti-fit when a branded or authorized custom voice is mandatory.

Listnr remains usable only as a conditional simple-studio route. Its product and pricing pages describe a browser flow from pasted text to voice selection and MP3/WAV download, and the pricing page presents annual credit packages with commercial rights. However, other first-party pages show different prices, while the API landing page and linked repository point to different hosts. Trial the editor if the requirement is simply to generate and download a voiceover; do not use Listnr as the price winner or API recommendation until checkout and endpoint behavior are reconciled.

Voice cloning changes the rights check

A technically available clone is not automatically an authorized one. ElevenLabs requires the user to confirm rights and consent for instant cloning and limits its professional process to the user's verified voice. Cartesia requires explicit consent for another person's recordings and prohibits impersonation and rights violations. Fish Audio places responsibility for input rights, privacy, disclosure, and commercial use on the user. Listnr's terms require verifiable consent and written permission for identifiable speakers.

Keep a consent record that names the speaker, permitted uses, channels, territories, duration, compensation if any, disclosure approach, and a revocation path. Recheck whether the selected plan permits commercial output and whether a voice can be private, shared, or used through an API. Ownership of an exported audio file does not replace publicity, privacy, trademark, labor, or contractual permission.

Cloning also creates an operational boundary. Test whether a vendor supports a reusable private voice, a per-request reference, or only a vendor-reviewed professional clone. Ask how reference audio and generated output may be used for service improvement, how deletion works, and whether an opt-out or enterprise agreement is required. If those answers are unclear, use a licensed stock voice until the rights path is resolved.

Compare cost and capacity correctly

Do not compare a monthly app subscription, a creator credit, an input character, an UTF-8 byte, and a generated second as though they were interchangeable. Build a small cost model from the actual script mix. Include repeated generations during editing, language and encoding differences, silence or SSML rules, failed requests, and any top-up or overage price. For long-form work, include the human cost of splitting, revising, and re-exporting projects.

Keep app and API budgets separate unless the vendor explicitly shares quota. ElevenLabs is notable for a shared workspace pool. Speechify separates Reader, Studio, and API products; Murf separates Studio from its API; Fish separates web credits from API bytes. A low headline price can disappear if the needed commercial license, output format, private clone, team access, or concurrency sits on another route.

For live systems, monthly allowance is only one capacity dimension. Measure end-to-end time to first playable audio, sustained throughput, simultaneous active contexts, reconnect behavior, rate-limit responses, and fallback handling. Cartesia publishes plan-level context limits; Fish ties concurrency to prepaid balance; Unreal's public capacity documentation is too old for a confident plan comparison. Buy from measured peaks and failure tolerance, not a vendor's isolated latency claim.

A practical trial sequence

1. Define one primary lane: reading, creator narration, business voiceover, real-time speech, cloning, or low-cost batch generation. Record any secondary lane so it does not quietly dictate the purchase.
2. Use the same test corpus across candidates: short dialogue, long narration, names, numbers, abbreviations, punctuation, and every production language. Include only voices whose commercial and cloning permissions are clear.
3. Complete the real output workflow. Import the source, revise pronunciation and pacing, regenerate one section, export the required format, and verify that timestamps, chapters, subtitles, or project sharing survive as needed.
4. For an API, run from the intended region with the intended streaming method. Capture latency percentiles, active concurrency, 429 and timeout behavior, retry cost, audio format, and long-job completion—not just one successful request.
5. Normalize the bill to the vendor's real unit and expected re-generation rate. Model the first paid tier, the likely steady-state tier, overages, annual commitment, and any separate app, API, seat, clone, or enterprise charge.
6. Finish with a rights and operations gate. Confirm commercial use, consent, disclosure, data handling, deletion, support, and a fallback voice or provider before production traffic or public distribution.

The decision is complete when one route passes the workflow, rights, capacity, and cost gates for the primary job. Keep ElevenLabs as the broad first trial when those needs are still mixed. Branch to a specialist as soon as the real constraint becomes reading continuity, business review, real-time contexts, cloning flexibility, or low-cost character volume.