@elizaos/shared 2.0.0-alpha.538 → 2.0.3-beta.2

package/local-inference/voice-models.js

@@ -0,0 +1,746 @@
+/**
+ * Sub-model versioning for voice components.
+ *
+ * The bundle manifest (`eliza-1.manifest.json`) ships the *current* set of
+ * files for one tier. This module ships the *history*: every voice
+ * sub-model id we publish, every semver version, the parent it succeeds,
+ * the per-version eval deltas vs. that parent, the GGUF assets, and the
+ * minimum bundle version each voice version is compatible with.
+ *
+ * The publish pipeline writes both this file AND the matching H3 in
+ * the sibling `voice-models/CHANGELOG.md`. The publish gate refuses to
+ * land one without the other. The runtime auto-update checker reads
+ * only this file.
+ *
+ * Spec: `.swarm/research/R5-versioning.md` §2.
+ */
+/**
+ * Reverse-chronological history per model id. Index 0 is the latest.
+ *
+ * The publish pipeline prepends a new version; never edit a published
+ * entry in place (sha + size are the audit trail).
+ *
+ * Asset hashes and revisions are pinned audit data. `missingAssets` records
+ * unavailable upstream artifacts explicitly instead of using provisional
+ * checksums.
+ */
+export const VOICE_MODEL_VERSIONS = [
+    {
+        id: "wakeword",
+        version: "0.3.0",
+        parentVersion: "0.2.0",
+        publishedToHfAt: "2026-06-19T18:05:39Z",
+        hfRepo: "elizaos/eliza-1",
+        hfRevision: "c544bb4c78a601a0da8372b9399dfe668fbadb1e",
+        ggufAssets: [
+            {
+                filename: "voice/wakeword/hey-eliza.melspec.gguf",
+                sha256: "98bd2d5e3cc09e416626cd1a6a758cb92bb8096766d25ecf57d4c99927db682d",
+                sizeBytes: 543584,
+                quant: "fp16",
+            },
+            {
+                filename: "voice/wakeword/hey-eliza.embedding.gguf",
+                sha256: "9cfcb0d9f1939c68cc9e63f5ac9e0f09b8e8568ee1085dbb50e7391e874a1dc5",
+                sizeBytes: 659904,
+                quant: "fp16",
+            },
+            {
+                filename: "voice/wakeword/hey-eliza.classifier.gguf",
+                sha256: "4502c92664b18d598753114f09925921ddd065d72871607c3a842fa70510a350",
+                sizeBytes: 315456,
+                quant: "fp16",
+            },
+        ],
+        evalDeltas: { netImprovement: true },
+        changelogEntry: "0.3.0 - real hey-eliza head (no-mel-rescale, ~98pct TA / ~4-7pct FA held-out)",
+        minBundleVersion: "0.0.0",
+    },
+    {
+        id: "turn-detector-intl",
+        version: "0.1.0",
+        publishedToHfAt: "2026-05-15T11:15:08Z",
+        hfRepo: "elizaos/eliza-1",
+        hfRevision: "e7ef6204cbede995cc1ff740ed448ce1b6fe93d2",
+        // K7: GGUF path live via LiveKitGgmlTurnDetector (J1.d). ONNX file
+        // stays on HF for one release per HF asset policy.
+        preferredBackend: "ggml",
+        deprecatedBackends: ["onnx"],
+        ggufAssets: [
+            {
+                filename: "voice/turn/intl/model_q8.onnx",
+                sha256: "af70f5b5e815f6baf11dad252fbc80400964c6589cea02115187139f6ccf9d66",
+                sizeBytes: 262031196,
+                quant: "onnx-int8",
+            },
+            {
+                filename: "voice/turn/intl/turn-detector-intl-q8.gguf",
+                sha256: "5dbcba3fb490217b10ec898003dd0905f9d81b8b7e24378029cff921ab7f9e79",
+                sizeBytes: 281016768,
+                quant: "q8_0",
+            },
+        ],
+        evalDeltas: { f1Delta: 0.09, netImprovement: true },
+        changelogEntry: "LiveKit v0.4.1-intl fine-tuned on multilingual EOU (OASST1 12-lang prompter, prefix-augmented), APOLLO-Mini, F1=0.9308",
+        minBundleVersion: "0.0.0",
+    },
+    {
+        id: "turn-detector",
+        version: "0.2.0",
+        parentVersion: "0.1.0",
+        publishedToHfAt: "2026-05-15T05:17:55Z",
+        hfRepo: "elizaos/eliza-1",
+        hfRevision: "20b291b5820937e8a1e1ca9f2927f5bc64aefe7e",
+        // K7: GGUF path live via LiveKitGgmlTurnDetector (J1.d). ONNX file
+        // stays on HF for one release per HF asset policy.
+        preferredBackend: "ggml",
+        deprecatedBackends: ["onnx"],
+        ggufAssets: [
+            {
+                filename: "voice/turn-detector/onnx/model_q8.onnx",
+                sha256: "52a132ed9c53fe41381cd97a800a9d36b7494d5ea608183de03adfc0723662f8",
+                sizeBytes: 37736316,
+                quant: "onnx-int8",
+            },
+            {
+                filename: "voice/turn-detector/onnx/turn-detector-en-q8.gguf",
+                sha256: "04bc18aeec5f59a94ae338aa66a48e204fc02785a6217d9055145f95d2192980",
+                sizeBytes: 41275296,
+                quant: "q8_0",
+            },
+        ],
+        evalDeltas: { f1Delta: 0.1411, netImprovement: true },
+        changelogEntry: "LiveKit v1.2.2-en fine-tuned on DailyDialog (prefix-augmented EOU corpus, APOLLO-Mini, F1=0.9811 vs 0.84 baseline)",
+        minBundleVersion: "0.0.0",
+    },
+    {
+        id: "omnivoice",
+        version: "0.2.0",
+        parentVersion: "0.1.0",
+        publishedToHfAt: "2026-05-15T04:50:24Z",
+        hfRepo: "elizaos/eliza-1",
+        hfRevision: "20b291b5820937e8a1e1ca9f2927f5bc64aefe7e",
+        // K7: OmniVoice is fully on the fork via libelizainference FFI (W3-3).
+        // No ONNX path ever existed for OmniVoice.
+        preferredBackend: "llama-server",
+        ggufAssets: [
+            {
+                filename: "voice/omnivoice/omnivoice-base-q4_k_m.gguf",
+                sha256: "4836ba5affcb83c04d08b3e96b6e48ac839947ee3864188739807b68e0e159c6",
+                sizeBytes: 407_485_216,
+                quant: "q4_k_m",
+            },
+            {
+                filename: "voice/omnivoice/omnivoice-tokenizer-q4_k_m.gguf",
+                sha256: "988fc32bc699bce361c44e9af8383be9811960441f293eb5ab4ec6bf6386378d",
+                sizeBytes: 252_474_112,
+                quant: "q4_k_m",
+            },
+            {
+                filename: "voice/omnivoice/omnivoice-base-q8_0.gguf",
+                sha256: "2882d887921798aea13d45236556bdf8012842ab6f8cd2690943eead6289f298",
+                sizeBytes: 656_395_008,
+                quant: "q8_0",
+            },
+            {
+                filename: "voice/omnivoice/presets/voice-preset-same.bin",
+                sha256: "efb3ab57f6e3884a2414a9cf9dcdb77a66b61bb43713df9ff1632e7539191be6",
+                sizeBytes: 716,
+                quant: "fp16",
+            },
+        ],
+        evalDeltas: { netImprovement: true },
+        changelogEntry: "OmniVoice ELZ2 v2 frozen-conditioning preset for the 'same' voice (Her-derivative)",
+        minBundleVersion: "0.0.0",
+    },
+    {
+        id: "speaker-encoder",
+        version: "0.2.0",
+        parentVersion: "0.1.0",
+        publishedToHfAt: "2026-05-19T00:00:00Z",
+        hfRepo: "elizaos/eliza-1",
+        hfRevision: "pending",
+        preferredBackend: "ffi",
+        deprecatedBackends: ["onnx"],
+        ggufAssets: [],
+        missingAssets: [
+            {
+                filename: "voice/speaker-encoder/wespeaker-resnet34-lm.gguf",
+                quant: "gguf-fp32",
+                expectedSizeBytes: 27_000_000,
+                reason: "missing-from-hf-repo",
+            },
+        ],
+        evalDeltas: { netImprovement: true },
+        changelogEntry: "0.2.0 — GGUF conversion via voice_speaker_to_gguf.py; ONNX file removed from HF.",
+        minBundleVersion: "0.0.0",
+    },
+    {
+        id: "speaker-encoder",
+        version: "0.1.0",
+        publishedToHfAt: "2026-05-15T07:15:30Z",
+        hfRepo: "elizaos/eliza-1",
+        hfRevision: "20b291b5820937e8a1e1ca9f2927f5bc64aefe7e",
+        deprecatedBackends: ["onnx"],
+        ggufAssets: [
+            {
+                filename: "voice/speaker-encoder/wespeaker-resnet34-lm.onnx",
+                sha256: "7bb2f06e9df17cdf1ef14ee8a15ab08ed28e8d0ef5054ee135741560df2ec068",
+                sizeBytes: 26_530_309,
+                quant: "onnx-fp16",
+            },
+        ],
+        evalDeltas: { netImprovement: true },
+        changelogEntry: "Initial release — WeSpeaker ResNet34-LM 256-dim ONNX (deprecated; GGUF in 0.2.0).",
+        minBundleVersion: "0.0.0",
+    },
+    {
+        id: "diarizer",
+        version: "0.2.0",
+        parentVersion: "0.1.0",
+        publishedToHfAt: "2026-05-19T00:00:00Z",
+        hfRepo: "elizaos/eliza-1",
+        hfRevision: "pending",
+        preferredBackend: "ffi",
+        deprecatedBackends: ["onnx"],
+        ggufAssets: [],
+        missingAssets: [
+            {
+                filename: "voice/diarizer/pyannote-segmentation-3.0.gguf",
+                quant: "gguf-fp32",
+                expectedSizeBytes: 6_000_000,
+                reason: "missing-from-hf-repo",
+            },
+        ],
+        evalDeltas: { netImprovement: true },
+        changelogEntry: "0.2.0 — GGUF conversion via voice_diarizer_to_gguf.py; ONNX files removed from HF.",
+        minBundleVersion: "0.0.0",
+    },
+    {
+        id: "diarizer",
+        version: "0.1.0",
+        publishedToHfAt: "2026-05-15T07:15:30Z",
+        hfRepo: "elizaos/eliza-1",
+        hfRevision: "20b291b5820937e8a1e1ca9f2927f5bc64aefe7e",
+        deprecatedBackends: ["onnx"],
+        ggufAssets: [
+            {
+                filename: "voice/diarizer/pyannote-segmentation-3.0-int8.onnx",
+                sha256: "465d0975bf70fbf14fb77c0589a5d346a9c07c2170345f529cf774678446db76",
+                sizeBytes: 1_542_304,
+                quant: "onnx-int8",
+            },
+            {
+                filename: "voice/diarizer/pyannote-segmentation-3.0-fp32.onnx",
+                sha256: "057ee564753071c0b09b5b611648b50ac188d50846bff5f01e9f7bbf1591ea25",
+                sizeBytes: 5_986_908,
+                quant: "onnx-fp16",
+            },
+        ],
+        evalDeltas: { netImprovement: true },
+        changelogEntry: "Initial release — Pyannote-segmentation-3.0 ONNX int8 (deprecated; GGUF in 0.2.0).",
+        minBundleVersion: "0.0.0",
+    },
+    {
+        id: "turn-detector",
+        version: "0.1.0",
+        publishedToHfAt: "2026-05-14T00:00:00Z",
+        hfRepo: "elizaos/eliza-1",
+        hfRevision: "20b291b5820937e8a1e1ca9f2927f5bc64aefe7e",
+        ggufAssets: [
+            {
+                filename: "voice/turn-detector/turn-detector-en-int8.onnx",
+                sha256: "fdd695a99bda01155fb0b5ce71d34cb9fd3902c62496db7a6c2c7bdeac310ac7",
+                sizeBytes: 65_712_276,
+                quant: "onnx-int8",
+            },
+            {
+                filename: "voice/turn-detector/turn-detector-intl-int8.onnx",
+                sha256: "bd2c30776882138a1d95a07faddc13756fe1a35bef6323505f1124fca349bc9c",
+                sizeBytes: 396_316_457,
+                quant: "onnx-int8",
+            },
+            {
+                filename: "voice/turn-detector/turnsense-fallback-int8.onnx",
+                sha256: "a423adf55f5f33cf4ee9e3fe73ec133d0106affae3aa14693417b4a1c79e2df8",
+                sizeBytes: 176_072_860,
+                quant: "onnx-int8",
+            },
+        ],
+        evalDeltas: { netImprovement: true },
+        changelogEntry: "Initial release — LiveKit turn-detector (135M + intl).",
+        minBundleVersion: "0.0.0",
+    },
+    {
+        id: "voice-emotion",
+        version: "0.3.0",
+        parentVersion: "0.2.0",
+        publishedToHfAt: "2026-05-19T00:00:00Z",
+        hfRepo: "elizaos/eliza-1",
+        hfRevision: "pending",
+        preferredBackend: "ffi",
+        deprecatedBackends: ["onnx"],
+        ggufAssets: [],
+        missingAssets: [
+            {
+                filename: "voice/voice-emotion/wav2small-msp-dim.gguf",
+                quant: "gguf-fp32",
+                expectedSizeBytes: 1_200_000,
+                reason: "missing-from-hf-repo",
+            },
+        ],
+        evalDeltas: { netImprovement: true },
+        changelogEntry: "0.3.0 — GGUF conversion via voice_emotion_to_gguf.py; ONNX files removed from HF.",
+        minBundleVersion: "0.0.0",
+    },
+    {
+        id: "voice-emotion",
+        version: "0.2.0",
+        publishedToHfAt: "2026-05-15T07:20:39Z",
+        hfRepo: "elizaos/eliza-1",
+        hfRevision: "20b291b5820937e8a1e1ca9f2927f5bc64aefe7e",
+        deprecatedBackends: ["onnx"],
+        ggufAssets: [
+            {
+                filename: "voice/emotion/wav2small-cls7-int8.onnx",
+                sha256: "cba2c4e49707ac20da8b1420814b80735f700e917905c46d8cb880b95d97c953",
+                sizeBytes: 524_750,
+                quant: "onnx-int8",
+            },
+            {
+                filename: "voice/voice-emotion/wav2small-msp-dim-int8.onnx",
+                sha256: "2fcde4aa2a6881b0e7407a3a706fab1889b69233139ee10b8669795b02b06efc",
+                sizeBytes: 516_877,
+                quant: "onnx-int8",
+            },
+            {
+                filename: "voice/voice-emotion/wav2small-msp-dim-fp32.onnx",
+                sha256: "3f5a8bf8f035132798b170c57ab61b90d52bb0cb0dd1ef95fd40a97f466f65f7",
+                sizeBytes: 1_211_917,
+                quant: "onnx-fp32",
+            },
+            {
+                filename: "voice/voice-emotion/wav2small-msp-dim-fp32.onnx.data",
+                sha256: "5a3a84e879c786317570b551ee6240294c8deded27856a52d872c03b12c63d01",
+                sizeBytes: 989_472,
+                quant: "onnx-fp32",
+            },
+        ],
+        evalDeltas: { f1Delta: 0.0042, netImprovement: true },
+        changelogEntry: "v0.2.0 — Wav2Small cls7 head shipped (macro-F1=0.355 > 0.35 gate). ONNX files deprecated; GGUF in 0.3.0.",
+        minBundleVersion: "0.0.0",
+    },
+    {
+        // J2 (2026-05-15): Kokoro GGUF asset for the fork-side inference path.
+        // The fork's `tools/kokoro/` subtree implements a standalone StyleTTS-2
+        // + iSTFTNet inference pipeline (LLM_ARCH_KOKORO arch + GGML graph +
+        // CPU iSTFT vocoder). The runtime selector (`pickKokoroRuntimeBackend`)
+        // defaults `KOKORO_BACKEND=fork` and POSTs to llama-server's
+        // /v1/audio/speech route; ONNX path remains as one-release deprecation.
+        //
+        // Quality gap: the from-scratch port runs at lower acoustic quality
+        // vs the ONNX baseline (the predictor convs + ResBlock decoder need
+        // a follow-up per-tensor weight-mapping pass). Documented in
+        // .swarm/impl/J2-kokoro-port-notes.md; ship continues per brief override.
+        //
+        // missingAssets carries the planned ladder; the GGUF push to the
+        // consolidated elizaos/eliza-1 repo lands once the full PyTorch
+        // checkpoint walks through convert_kokoro_pth_to_gguf.py with the full
+        // _PTH_KEY_RULES map (Q3..Q8 ladder via gguf_kokoro_apply.py from W3-1).
+        id: "kokoro",
+        version: "0.3.0",
+        parentVersion: "0.2.0",
+        publishedToHfAt: "2026-05-15T05:00:00Z",
+        hfRepo: "elizaos/eliza-1",
+        hfRevision: "4b8809b197aa90ae486f83c1e0a5dc7effb6b285",
+        // K7: runtime defaults to KOKORO_BACKEND=fork (pick-runtime.ts). ONNX
+        // path is one-release deprecation runway (KOKORO_BACKEND=onnx env only).
+        // GGUF conversion pending (K4 scope; missingAssets carries the target).
+        preferredBackend: "llama-server",
+        deprecatedBackends: ["onnx"],
+        ggufAssets: [
+            {
+                filename: "voice/kokoro/voices/af_sam.bin",
+                sha256: "6874670865ce984a5400afc87176706c5ed88671999c59ed0dff5dcde664277b",
+                sizeBytes: 522_240,
+                quant: "fp16",
+            },
+        ],
+        missingAssets: [
+            {
+                filename: "voice/kokoro/kokoro-v1.0-q4_k_m.gguf",
+                quant: "q4_k_m",
+                expectedSizeBytes: 60_000_000,
+                reason: "missing-from-local-staging",
+            },
+        ],
+        evalDeltas: {
+            // Same af_sam embedding ships as in 0.2.0. The version bump tracks
+            // the runtime path move (ONNX → fork llama-server) not an embedding
+            // change. netImprovement=false until the acoustic-quality gap closes.
+            netImprovement: false,
+        },
+        changelogEntry: "0.3.0 — J2: fork-side Kokoro inference path (LLM_ARCH_KOKORO graph + tools/kokoro/). Runtime selector defaults to KOKORO_BACKEND=fork → llama-server /v1/audio/speech. ONNX path retained for one release. Quality gap vs ONNX baseline documented; compute-gated follow-up for full per-tensor weight mapping.",
+        minBundleVersion: "0.0.0",
+    },
+    {
+        id: "kokoro",
+        version: "0.2.0",
+        parentVersion: "0.1.0",
+        publishedToHfAt: "2026-05-15T11:27:44Z",
+        hfRepo: "elizaos/eliza-1",
+        hfRevision: "e7ef6204cbede995cc1ff740ed448ce1b6fe93d2",
+        ggufAssets: [
+            {
+                // af_sam — mel-fit voice clone anchor=0.2; shipped per user override.
+                // evalGatePass=false: WER=1.0, UTMOS=2.32, SpkSim=0.15 (all fail gates).
+                // Production sam voice is OmniVoice preset; this is Kokoro eval slot.
+                filename: "voice/kokoro/voices/af_sam.bin",
+                sha256: "6874670865ce984a5400afc87176706c5ed88671999c59ed0dff5dcde664277b",
+                sizeBytes: 522_240,
+                quant: "fp16",
+            },
+        ],
+        evalDeltas: {
+            // Negative delta vs baseline af_bella: UTMOS -1.88, WER +1.0, SpkSim +0.23.
+            // netImprovement=false — quality regression. Auto-updater will not recommend swap.
+            mosDelta: -1.88,
+            werDelta: 1.0,
+            netImprovement: false,
+        },
+        changelogEntry: "0.2.0 — af_sam shipped per user override (I2); eval WER=1.0/UTMOS=2.32/SpkSim=0.15, gate=FAIL. OmniVoice preset remains production sam voice. compute-gated: needs ≥3h corpus for real quality.",
+        minBundleVersion: "0.0.0",
+    },
+    {
+        id: "kokoro",
+        version: "0.1.0",
+        publishedToHfAt: "2026-05-14T00:00:00Z",
+        hfRepo: "elizaos/eliza-1",
+        hfRevision: "20b291b5820937e8a1e1ca9f2927f5bc64aefe7e",
+        deprecatedBackends: ["onnx"],
+        ggufAssets: [
+            {
+                // kokoro-v1.0-q4.onnx removed from HF; GGUF via 0.3.0 fork path.
+                filename: "voice/kokoro/kokoro-v1.0-q4.onnx",
+                sha256: "04cf570cf9c4153694f76347ed4b9a48c1b59ff1de0999e6605d123966b197c7",
+                sizeBytes: 305_215_966,
+                quant: "onnx-int8",
+            },
+            {
+                filename: "voice/kokoro/voices/af_bella.bin",
+                sha256: "f69d836209b78eb8c66e75e3cda491e26ea838a3674257e9d4e5703cbaf55c8b",
+                sizeBytes: 522_240,
+                quant: "fp16",
+            },
+        ],
+        evalDeltas: { netImprovement: true },
+        changelogEntry: "Initial release — kokoro 82M voice-embedding af_bella preset (ONNX deprecated; fork GGUF path in 0.3.0).",
+        minBundleVersion: "0.0.0",
+    },
+    {
+        id: "omnivoice",
+        version: "0.1.0",
+        publishedToHfAt: "2026-05-14T00:00:00Z",
+        hfRepo: "elizaos/eliza-1",
+        hfRevision: "20b291b5820937e8a1e1ca9f2927f5bc64aefe7e",
+        ggufAssets: [],
+        missingAssets: [
+            {
+                filename: "voice/omnivoice/omnivoice-base-q4_k_m.gguf",
+                quant: "q4_k_m",
+                expectedSizeBytes: 388_000_000,
+                reason: "missing-from-local-staging",
+            },
+            {
+                filename: "voice/omnivoice/omnivoice-tokenizer-q4_k_m.gguf",
+                quant: "q4_k_m",
+                expectedSizeBytes: 51_200_000,
+                reason: "missing-from-local-staging",
+            },
+            {
+                filename: "voice/omnivoice/omnivoice-base-q8_0.gguf",
+                quant: "q8_0",
+                expectedSizeBytes: 620_000_000,
+                reason: "missing-from-local-staging",
+            },
+            {
+                filename: "voice/omnivoice/presets/voice-preset-same.bin",
+                quant: "fp16",
+                expectedSizeBytes: 8_192,
+                reason: "missing-from-local-staging",
+            },
+        ],
+        evalDeltas: { netImprovement: true },
+        changelogEntry: "Initial release — OmniVoice frozen-conditioning sam.",
+        minBundleVersion: "0.0.0",
+    },
+    {
+        id: "vad",
+        version: "0.2.0",
+        parentVersion: "0.1.0",
+        publishedToHfAt: "2026-05-19T00:00:00Z",
+        hfRepo: "elizaos/eliza-1",
+        hfRevision: "pending",
+        // K7: VAD is fully native through silero-vad-cpp/libsilero_vad.
+        // ONNX file removed from HF in this release.
+        preferredBackend: "ffi",
+        ggufAssets: [
+            {
+                filename: "voice/vad/silero-vad-v5.gguf",
+                sha256: "d348cd6d87ea53dcd3e6680698c88be326082e27dae899adef653d090bee4995",
+                sizeBytes: 620_736,
+                quant: "gguf-fp32",
+            },
+        ],
+        evalDeltas: { netImprovement: true },
+        changelogEntry: "0.2.0 — ONNX file removed from HF; GGUF-only release.",
+        minBundleVersion: "0.0.0",
+    },
+    {
+        id: "vad",
+        version: "0.1.0",
+        publishedToHfAt: "2026-05-15T07:15:30Z",
+        hfRepo: "elizaos/eliza-1",
+        hfRevision: "20b291b5820937e8a1e1ca9f2927f5bc64aefe7e",
+        // K7: VAD is fully native through silero-vad-cpp/libsilero_vad.
+        // vad.ts imports SileroVadGgml only; onnxruntime-node NOT imported.
+        // ONNX file deprecated; removed from HF in 0.2.0.
+        preferredBackend: "ffi",
+        deprecatedBackends: ["onnx"],
+        ggufAssets: [
+            {
+                filename: "voice/vad/silero-vad-v5.gguf",
+                sha256: "d348cd6d87ea53dcd3e6680698c88be326082e27dae899adef653d090bee4995",
+                sizeBytes: 620_736,
+                quant: "gguf-fp32",
+            },
+            {
+                filename: "voice/vad/silero-vad-int8.onnx",
+                sha256: "90b026c95f054d59d7bf79387b0ed93c8950f35a4d8b741cd78d4bb23a7d2776",
+                sizeBytes: 639_383,
+                quant: "onnx-int8",
+            },
+        ],
+        evalDeltas: { netImprovement: true },
+        changelogEntry: "Initial release — Silero VAD v5 GGUF (ONNX file deprecated; removed in 0.2.0).",
+        minBundleVersion: "0.0.0",
+    },
+    {
+        id: "wakeword",
+        version: "0.2.0",
+        parentVersion: "0.1.0",
+        publishedToHfAt: "2026-05-19T00:00:00Z",
+        hfRepo: "elizaos/eliza-1",
+        hfRevision: "pending",
+        // K7: wake-word is fully on fork FFI. ONNX files removed from HF in
+        // this release; combined GGUF (mel filterbank + embedding + heads) is
+        // the canonical asset going forward.
+        preferredBackend: "ffi",
+        ggufAssets: [],
+        missingAssets: [
+            {
+                filename: "voice/wakeword/openwakeword.gguf",
+                quant: "gguf-fp32",
+                expectedSizeBytes: 3_000_000,
+                reason: "missing-from-hf-repo",
+            },
+        ],
+        evalDeltas: { netImprovement: true },
+        changelogEntry: "0.2.0 — ONNX files removed from HF; combined openwakeword.gguf target.",
+        minBundleVersion: "0.0.0",
+    },
+    {
+        id: "wakeword",
+        version: "0.1.0",
+        publishedToHfAt: "2026-05-15T07:15:30Z",
+        hfRepo: "elizaos/eliza-1",
+        hfRevision: "20b291b5820937e8a1e1ca9f2927f5bc64aefe7e",
+        // K7: wake-word is fully on fork FFI (eliza_inference_wakeword_* in
+        // libelizainference). wake-word.ts uses GgmlWakeWordModel only;
+        // onnxruntime-node NOT imported. ONNX files deprecated; removed in 0.2.0.
+        preferredBackend: "ffi",
+        deprecatedBackends: ["onnx"],
+        ggufAssets: [
+            {
+                filename: "voice/wakeword/melspectrogram.onnx",
+                sha256: "ba2b0e0f8b7b875369a2c89cb13360ff53bac436f2895cced9f479fa65eb176f",
+                sizeBytes: 1_087_958,
+                quant: "onnx-fp32",
+            },
+            {
+                filename: "voice/wakeword/embedding_model.onnx",
+                sha256: "70d164290c1d095d1d4ee149bc5e00543250a7316b59f31d056cff7bd3075c1f",
+                sizeBytes: 1_326_578,
+                quant: "onnx-fp32",
+            },
+            {
+                filename: "voice/wakeword/hey-eliza-int8.onnx",
+                sha256: "e565952901cd4203baacef7cb8700891c9bee4e6f42fc9bc0aa03b9c39a2da92",
+                sizeBytes: 630_032,
+                quant: "onnx-int8",
+            },
+        ],
+        evalDeltas: { netImprovement: true },
+        changelogEntry: "Initial release — hey-eliza wake-word head ONNX (deprecated; removed in 0.2.0).",
+        minBundleVersion: "0.0.0",
+    },
+    {
+        id: "embedding",
+        version: "0.1.0",
+        publishedToHfAt: "2026-05-15T07:15:30Z",
+        hfRepo: "elizaos/eliza-1",
+        hfRevision: "20b291b5820937e8a1e1ca9f2927f5bc64aefe7e",
+        ggufAssets: [
+            {
+                filename: "voice/embedding/eliza-1-embedding-q8_0.gguf",
+                sha256: "06507c7b42688469c4e7298b0a1e16deff06caf291cf0a5b278c308249c3e439",
+                sizeBytes: 639_150_592,
+                quant: "q8_0",
+            },
+        ],
+        evalDeltas: { netImprovement: true },
+        changelogEntry: "Initial release — Eliza-1 BPE-vocab embedding tier.",
+        minBundleVersion: "0.0.0",
+    },
+    {
+        id: "asr",
+        version: "0.1.0",
+        publishedToHfAt: "2026-05-14T00:00:00Z",
+        hfRepo: "elizaos/eliza-1",
+        hfRevision: "20b291b5820937e8a1e1ca9f2927f5bc64aefe7e",
+        ggufAssets: [
+            {
+                filename: "voice/asr/eliza-1-asr-q8_0.gguf",
+                sha256: "58e22d0532d4eacaf034cfac17a6fed159f37c41390c710186783be439d1fc57",
+                sizeBytes: 2_165_034_944,
+                quant: "q8_0",
+            },
+            {
+                filename: "voice/asr/eliza-1-asr-mmproj.gguf",
+                sha256: "46c1d533af3f354ceb37ce855dbceff7da7fa7cf1e6a523df3b13440bd164c0d",
+                sizeBytes: 355_709_344,
+                quant: "q8_0",
+            },
+        ],
+        evalDeltas: { netImprovement: true },
+        changelogEntry: "Initial release — Qwen3-ASR streaming transcriber Q8_0.",
+        minBundleVersion: "0.0.0",
+    },
+    {
+        id: "asr",
+        version: "0.2.0",
+        publishedToHfAt: "2026-05-15T00:00:00Z",
+        hfRepo: "elizaos/eliza-1",
+        hfRevision: "50cffb075ae3c24a4b0cd3a8ccdfaa92506f70d4",
+        ggufAssets: [
+            {
+                filename: "voice/asr/eliza-1-asr-q3_k_m.gguf",
+                sha256: "80c046bf44cf52699bcef0f5a10d9774a0415c127432a0d0f428e6f1634d7f12",
+                sizeBytes: 1_073_237_952,
+                quant: "q3_k_m",
+            },
+            {
+                filename: "voice/asr/eliza-1-asr-q4_k_m.gguf",
+                sha256: "de11f7110e6faddc277088262cffe8adc3228220a1cfd133c49420bb8bd3c3c5",
+                sizeBytes: 1_282_435_008,
+                quant: "q4_k_m",
+            },
+            {
+                filename: "voice/asr/eliza-1-asr-q5_k_m.gguf",
+                sha256: "dcffc861bc968202a6ca1107f3335c3dfcd89d6945ca87569596d6a9f551207a",
+                sizeBytes: 1_471_801_280,
+                quant: "q5_k_m",
+            },
+            {
+                filename: "voice/asr/eliza-1-asr-q6_k.gguf",
+                sha256: "b0f6f2d610dfcc95c4147271855d1aadf78d298d193666834912eb600c5eaf0b",
+                sizeBytes: 1_673_002_944,
+                quant: "q6_k",
+            },
+            {
+                filename: "voice/asr/eliza-1-asr-mmproj.gguf",
+                sha256: "46c1d533af3f354ceb37ce855dbceff7da7fa7cf1e6a523df3b13440bd164c0d",
+                sizeBytes: 355_709_344,
+                quant: "q8_0",
+            },
+        ],
+        evalDeltas: { netImprovement: true },
+        changelogEntry: "K-quant ladder: Q3_K_M, Q4_K_M, Q5_K_M, Q6_K added for Qwen3-ASR-1.7B. Enables memory-tier selection. mmproj stays Q8_0 per R8 §3.6.",
+        minBundleVersion: "0.0.0",
+    },
+];
+/**
+ * Strict semver compare. Returns -1 if a < b, 0 if equal, 1 if a > b.
+ * Pre-release ids are compared lexically per semver 2.0.0 §11. Returns
+ * null when either argument is not a valid semver.
+ */
+export function compareVoiceModelSemver(a, b) {
+    const parse = (s) => {
+        const m = /^(\d+)\.(\d+)\.(\d+)(?:-([A-Za-z0-9.-]+))?$/.exec(s);
+        if (!m)
+            return null;
+        return {
+            core: [Number(m[1]), Number(m[2]), Number(m[3])],
+            pre: m[4] ? m[4].split(".") : [],
+        };
+    };
+    const pa = parse(a);
+    const pb = parse(b);
+    if (!pa || !pb)
+        return null;
+    for (let i = 0; i < 3; i++) {
+        if (pa.core[i] !== pb.core[i])
+            return pa.core[i] < pb.core[i] ? -1 : 1;
+    }
+    // Per semver 2.0.0 §11: a version with a pre-release tag has lower
+    // precedence than the same version without.
+    if (pa.pre.length === 0 && pb.pre.length === 0)
+        return 0;
+    if (pa.pre.length === 0)
+        return 1;
+    if (pb.pre.length === 0)
+        return -1;
+    const n = Math.max(pa.pre.length, pb.pre.length);
+    for (let i = 0; i < n; i++) {
+        const ai = pa.pre[i];
+        const bi = pb.pre[i];
+        if (ai === undefined)
+            return -1;
+        if (bi === undefined)
+            return 1;
+        const aNum = /^\d+$/.test(ai) ? Number(ai) : null;
+        const bNum = /^\d+$/.test(bi) ? Number(bi) : null;
+        if (aNum !== null && bNum !== null) {
+            if (aNum !== bNum)
+                return aNum < bNum ? -1 : 1;
+        }
+        else if (aNum !== null) {
+            return -1;
+        }
+        else if (bNum !== null) {
+            return 1;
+        }
+        else if (ai !== bi) {
+            return ai < bi ? -1 : 1;
+        }
+    }
+    return 0;
+}
+/** Return all versions for a given model id, latest first. */
+export function versionsFor(id) {
+    return VOICE_MODEL_VERSIONS.filter((v) => v.id === id).sort((a, b) => {
+        const cmp = compareVoiceModelSemver(a.version, b.version);
+        if (cmp === null)
+            return 0;
+        return cmp === 1 ? -1 : cmp === -1 ? 1 : 0;
+    });
+}
+/** Latest known version for the given id, or undefined if none. */
+export function latestVoiceModelVersion(id) {
+    return versionsFor(id)[0];
+}
+/** Lookup by id + exact version. */
+export function findVoiceModelVersion(id, version) {
+    return VOICE_MODEL_VERSIONS.find((v) => v.id === id && v.version === version);
+}
+//# sourceMappingURL=voice-models.js.map