A GitHub repository can document several speech model generations while licensing their weights differently. Fish Speech is a concrete reason to save a model-license manifest before any production deployment.
Start with the exact checkpoint
Do not treat a project name as a version. Record a model family, commit ID, weight filename, checksum and the license that accompanies that checkpoint. Fish Audio documentation distinguishes at least one research-licensed model from materials released under a community agreement. Feature claims from one checkpoint do not automatically transfer to another. A model card for a current release also does not prove that an older build was entitled to the same commercial rights.
Classify the activity correctly
A researcher running evaluation on local hardware, a content studio selling edited narration and a platform allowing consumers to generate voices are different use cases. The community license defines commercial eligibility and a special hosted-service restriction; research-licensed weights have their own limitations. Audit who controls the synthesis request and whether customers receive model access. If a producer distributes code or weights, additional notices may apply compared with sharing only audio output.
Measure performance independently
Repository demos may illustrate multilingual or multi-character capabilities and vendor-supplied hardware results. They are not independent proof that a model meets a reader’s GPU budget, latency goal or speaker similarity threshold. Build a fixed script set, disclose hardware and inference settings, and recruit native listeners for pronunciation checks. Do not recreate celebrity voices for benchmarking; use consenting speakers or prelicensed voices.
Plan maintenance and provenance
A deployment manifest should include dependency versions, third-party licenses, source-checkpoint hashes, model license acceptance date, and a process for receiving updates. Restrict a public API behind authentication, moderate requested content and log only the safety metadata needed. The same code running a private demo can become a contract-breaking hosted service if opened to self-serve customers without the required agreement.
Alternatives and handoff
A managed voice API can transfer infrastructure operation to a provider, but introduces plan entitlements, region and data-retention questions. A permissively licensed local engine may be simpler for narrow system notifications but have lower expressive control. The general LLM/model registry belongs to AI02; this AI05 article is confined to the speech-specific production and rights consequences.
Evidence, availability and commercial-rights boundary
Research cutoff: 25 October 2026. Provider documentation is evidence of its published interface or policy, not independent product testing. No performance score, hands-on trial, customer consent, territory-wide entitlement or legal clearance is claimed. Confirm the exact version, selected region, voice rights, data flow, account limits and current service terms before deployment. Unknown model release dates, account prices and plan limits are deliberately left unasserted; monetary claims belong to the separately dated AI10 pricing ledger. Where documentation is undated or dynamic, the public page is not treated as an archival snapshot of every earlier tier.