There are three useful approaches, and each puts the work in a different place.
1. **Record a person.** A person reads the final script into a phone, recorder, or computer, then trims mistakes and exports a delivery file. This gives the speaker direct control over emphasis, warmth, pace, and how a local name should sound. It also makes later wording changes expensive: a clean revision may require a pickup that matches the original room, microphone, voice, and energy.
2. **Generate text-to-speech.** A service renders the script with a stock or authorized synthetic voice. Wording changes are quick, and the same script can be regenerated. Names, acronyms, dates, and dialogue can expose pronunciation or phrasing problems. Some systems offer pronunciation controls, but those controls can be model-specific.
3. **Use a mixed workflow.** Record the parts where human presence matters, then use an editor for cleanup, assembly, music, or small synthesized corrections. Another version starts with TTS as a timing draft, replaces it with a human performance, and uses the draft only as a pacing guide. This adds handoffs but can separate the strengths of performance, revision, and finishing.
### What the checked sources establish
- Audacity's official manual documents recording, editing, and export to WAV or MP3, and recommends making a WAV or AIFF safety copy before editing. That supports it as one possible local recording-and-finishing route; it does not establish how good a given microphone, room, or performance will sound. [Source V1]
- ElevenLabs documents controls for pronunciation and text normalization. Its current guidance explicitly treats numbers, dates, abbreviations, and alphanumeric strings as edge cases, and its pronunciation documentation says behavior depends on the model and voice. The vendor's pricing page describes free and paid tiers, credit-based usage, commercial licensing on paid tiers, and higher-quality output on some tiers. Those are current plan claims, not a quote for this project. [Sources V2, V3]
- Descript's current terms say AI-voice training requires a consenting speaker, prohibit unauthorized voice imitation, and state that input and output are user content owned by the user to the extent protectable by law. The same terms warn that generated output may not be unique and that copyright status may be unclear. The terms also describe storage and human/vendor access for certain voice-training and quality-assurance uses. [Source V4]
### My comparative judgment, not a product test
| Criterion | Recorded person | Text-to-speech | Mixed workflow |
|---|---|---|---|
| Intelligibility | Can be excellent, but room noise, microphone technique, pace, and diction matter. | Can be even and clean, but a fluent voice can still say a critical detail incorrectly. | Can keep a human performance while using editing to remove distractions and tighten pace. |
| Pronunciation | The speaker can rehearse a name or acronym and respond to direction. | Needs explicit checking; a dictionary, phonetic spelling, or rewritten phrase may help but must be heard, not assumed. | A difficult line can be re-recorded or replaced, though a synthetic patch may sound discontinuous. |
| Revision effort | High when revisions require matched pickups or a full rerecord. | Low for text changes, but each regeneration still needs a listening pass. | Medium: targeted changes are possible, but the edit may become harder to manage. |
| Privacy | A fully local recorder/editor can minimize uploads. Consent is still needed from participants. | The script and possibly voice data go to a service; account settings and current terms need review. | Depends on which stages stay local and which use cloud services. |
| Usable exports | Lossless WAV is useful as a master; MP3 is useful for lightweight sharing. | Confirm the exact format, sample rate, watermark, and download rights for the chosen plan. | Keep an editable project plus a lossless master and a compressed listening copy. |
| Rights | Clearer when the script, performance, music, and ambience are original or licensed and the speaker consents in writing. | Check voice authorization, output terms, commercial-use conditions, and whether the output can be copyrighted in the relevant jurisdiction. | Rights inherit every component, so keep a simple asset and consent log. |
| Cost | Equipment, room, and staff time may dominate even when software is free. | Subscription or credits plus regeneration time may dominate. | Often pays for both recording time and editing or service access. |
This table is a set of hypotheses for a small community project. It is not evidence that one route will be clearer, faster, safer, or cheaper in the team's actual setting.
### A small, fair comparison
Use the same final script and the same delivery target for three arms: (A) a person recorded and edited locally, (B) stock-voice TTS, and (C) the human recording with only cleanup plus one deliberately corrected line. Do not clone a participant's voice for this first comparison. Export a WAV master and a normalized MP3 listening copy when the workflow permits it.
Use this approximately 90-second diagnostic passage in every arm:
> “Is that toaster finished?” Maya asked. “Not yet,” said Xochitl Quetzalli, “but bring it to the Riverside Repair Café on Saturday, October 17, from 10:30 a.m. to 1:45 p.m.” Volunteers will help neighbors examine lamps, jackets, bicycles, and small appliances. The HVAC team will also show how a filter fits, but we cannot accept refrigerators, microwaves, or anything leaking fuel. Admission is free. If you can, bring the broken item, its power cord, any missing screws, and a note describing what happened. For example: “The lamp flickers after 3 minutes,” or “The radio stopped tuning at 98.7 FM.” Children are welcome with an adult. We will have coffee, tea, and a quiet table for people who prefer written instructions. Repairs are not guaranteed, and you remain responsible for deciding whether an item is safe to use. Questions? Call 312-555-0148 or email fix-it@example.org. “Even if we cannot repair it,” Maya said, “we can learn how it works.” Please say Xochitl Quetzalli, HVAC, 98.7 FM, and the date and phone number exactly as written.
Before recording, replace the fictional date, phone number, email address, and venue with approved public details. Keep the unusual name only if the named person approves both the spelling and spoken form; otherwise use another consented diagnostic name.
Have three listeners, including one person who did not see the script, score each output without seeing the workflow label:
- write down the date, times, phone number, acronym, frequency, and unusual name they heard;
- mark any sentence that required replay;
- rate pace and warmth on a five-point scale;
- record production minutes, correction minutes, number of regenerations or pickups, uploads made, plan used, and final file formats;
- note whether a correction changed anything outside the target line.
The comparison should retain the raw recording or original generation, the final exports, the exact service settings, the approved script, and a one-page rights/privacy note. It should report observations for these three artifacts only, without generalizing from one voice or one room to an entire approach.
### Still uncertain
- Actual intelligibility, naturalness, name pronunciation, and edit continuity until the three outputs are heard.
- The exact current price and permitted use for the account and plan that would be used at experiment time.
- Whether a service retains uploaded scripts or audio, for how long, and under which account controls; this must be checked against the selected service, plan, region, and settings immediately before use.
- Whether generated speech is protectable by copyright, and what releases the organization requires. That is jurisdiction- and policy-specific legal work.
- Whether the venue, participant voices, music, sound effects, and contact details have all necessary permissions.
### Limitations and next experiment
This run reviewed documentation; it did not test voices, microphones, rooms, exports, account controls, or billing. The next experiment is the blinded three-arm comparison above, followed by a terms-and-settings check for the exact accounts before uploading any nonpublic script or voice recording.