# Encoder Profiles This page documents encoder profile choices for `github.com/hunydev/g729`. Profiles always emit normal 10-byte G.729 frames. None of the profiles is an ITU encoder byte-exact or certification claim. ## Product Default `NewEncoder()` and `NewStreamingEncoder()` use `EncoderProfileCore`. This is the product default for the v0.1.x line. Core is selected by listening quality, not PESQ alone. A diagnostic PESQ-led profile can score closer to the local `bcg729` black-box anchor on some private samples, but recent blind tests preferred Core because the PESQ-led candidate sounded slightly more muffled. Core keeps the public frame shape and decoder compatibility: - 80 int16 samples at 8 kHz in, 10 packed bytes out. - RTP payload type 18 compatible payload shape. - `annexb=no`; no SID/CNG/DTX. - Zero-allocation steady-state frame encode path. ## Core Fast `EncoderProfileCoreFast` is an explicit opt-in for low-latency or high-density deployments. It keeps the normal 10-byte G.729 payload shape and the same public decoder compatibility as Core, but reduces selected encoder search precision/budget: - The focused fixed-codebook fourth-loop budget is reduced from the Core frame-level 180-entry cap to 60 entries per frame. - Subframe 0 is capped at 30 entries so it cannot consume the whole fast frame budget. - Gain preselect uses the standard-width center solve instead of Core's wider preselect-center precision. - The wider gain predictor arithmetic is still retained; disabling it was tested and caused a large Pages sample quality regression. Core Fast is not the product default, not a decoder exact path, and not an ITU encoder byte-exact claim. Treat it as a throughput/latency trade-off profile: run local listening and regression checks before using it for production streams that are quality-sensitive. ## Core Algorithm Notes The Core open-loop path follows Annex A's raw-correlation per-range maxima before the normalized three-range merge, and range-3 override checks every lower-range submultiple with the Core `11/10` lift instead of only the current pairwise winner. Core closed-loop pitch refinement also evaluates the encodable P1/P2 fractional boundary codepoints rather than silently restricting the search to only the three fractions around the integer winner. Its fixed-codebook path uses the K3=0.4 focused threshold search from §3.8.1 with a 180-entry frame cap across the two subframes; subframe 0 is conservatively capped at 90 so it cannot consume the whole frame budget. Its LSP VQ path uses the sequential Annex A search rather than the broader diagnostic second-stage search. Its gain predictor keeps the wider int32 §3.9.1 math path and the Annex A GA/GB preselection search. ## Diagnostic Profiles Diagnostic profiles are intended for regression, PESQ experiments, and listening comparisons. They are not conformance claims and should not be used as stronger standards evidence than Core. `EncoderProfileQualityPESQ` keeps the broader historical Quality heuristic surface disabled, and enables native reconstructed-gain residual search, gain clip repair, and fixed-codebook residual reranking. It has strong PESQ NB scores on the current private sample set, but blind listening found it slightly more muffled than Core, so it is not the default. `EncoderProfileQuality` remains available as the older broad quality-heuristic profile for diagnostics. Quality uses the sequential Annex A LSP VQ path by default; the broader second-stage LSP search remains available only as an internal diagnostic knob. `EncoderProfileQualityAnnexALSP` is kept as an explicit alias for the sequential Annex A LSP path in listening diagnostics. `EncoderProfileQualityClean` is a listening-diagnostic profile for comparing a smoother, `bcg729`-like candidate against the default Quality profile. It keeps the standard 10-byte payload shape, disables normalized closed-loop pitch reranking, and uses stricter, high-residual-aware decoder-in-loop gain MSE repair. `EncoderProfileQualityCleanSNR` keeps the same pitch policy but uses the older high-SNR gain-repair preference for clarity A/B tests. `EncoderProfileQualityCleanSmooth` lowers the clean repair threshold and biases more strongly toward high-residual reduction for bitstream-level smoothing diagnostics. `EncoderProfileQualityCleanVoiced` keeps the clean pitch policy while allowing decoder-in-loop gain repair to prefer slightly higher adaptive gain when the objective-score tradeoff remains bounded. `EncoderProfileQualityCleanDegrit` keeps the clean pitch policy while allowing gain repair to prefer lower fixed-codebook gain correction when adaptive gain is not reduced. `EncoderProfileQualityCleanHarmonic` keeps the clean pitch policy while allowing voiced gain repair to trade bounded score loss for higher adaptive gain with lower fixed-codebook correction. `EncoderProfileQualityCleanHarmonicStrong` pushes that same gain-balance tradeoff harder for grit-vs-muffling A/B tests. `EncoderProfileQualityCleanHarmonicDeep` pushes the same gain-balance tradeoff beyond the strong candidate to locate the grit-vs-muffling boundary. `EncoderProfileQualityCleanFCBRerank` keeps the clean pitch policy while reranking a small fixed-codebook candidate set with decoder-in-loop residual scoring for grit/noise listening diagnostics. `EncoderProfileCoreClipRepair` is a listening-diagnostic variant of Core. It keeps the Core LSP/FCB/gain-preselect policy and only adds decoder-in-loop gain clip repair with a lower pre-clip threshold. Use it for A/B tests of the Core sound with a minimal clipping safety valve, not as a conformance claim. ## Selection Guidance Use `NewEncoder()` or `NewEncoderWithProfile(EncoderProfileCore)` for normal applications. Use `NewEncoderWithProfile(EncoderProfileCoreFast)` only when channel density or latency is more important than preserving the default Core search result. Use diagnostic profiles only when you are explicitly running the private quality/PESQ/blind-listening test workflows described in [validation.md](validation.md).