@elizaos/plugin-local-inference 2.0.0-beta.1 → 2.0.3-beta.2

package/src/services/voice/voice-preset-format.ts

@@ -0,0 +1,700 @@
+/**
+ * Binary format for `cache/voice-preset-*.bin`.
+ *
+ * Two versions are supported:
+ *
+ *   v1 (`magic='ELZ1', version=1`) — legacy two-section layout used by the
+ *   initial Kokoro-style placeholder. Carries a Float32 speaker embedding +
+ *   a phrase-cache seed list. Still read for back-compat (older bundles only
+ *   contain v1).
+ *
+ *   v2 (`magic='ELZ1', version=2`) — superset adopted for the OmniVoice
+ *   freeze. Adds three OmniVoice-specific sections that the v1 layout had
+ *   no room for: pre-encoded `ref_audio_tokens` (int32, shape
+ *   `[K, ref_T]`), a UTF-8 `ref_text` transcript of the reference clip, and
+ *   a closed-vocabulary `instruct` string (the resolved VoiceDesign
+ *   attributes). v2 readers handle v1 files transparently (the new sections
+ *   default to empty). A v1 reader applied to a v2 file fails fast on
+ *   `truncated-header` because the v2 header is larger.
+ *
+ * Layout (little-endian throughout):
+ *
+ *   v1 header (24 bytes):
+ *     +0   4 bytes  magic 'ELZ1' (0x315A4C45)
+ *     +4   4 bytes  format version (uint32)                — 1
+ *     +8   4 bytes  speaker embedding offset (uint32)
+ *     +12  4 bytes  speaker embedding byte length (uint32)
+ *     +16  4 bytes  phrase cache seed offset (uint32)
+ *     +20  4 bytes  phrase cache seed byte length (uint32)
+ *
+ *   v2 header (64 bytes — additive, all section descriptors are
+ *               `(offset:uint32, length:uint32)` pairs):
+ *     +0   4 bytes  magic 'ELZ1' (0x315A4C45)
+ *     +4   4 bytes  format version (uint32)                — 2
+ *     +8   4 bytes  speaker embedding offset
+ *     +12  4 bytes  speaker embedding byte length
+ *     +16  4 bytes  phrase cache seed offset
+ *     +20  4 bytes  phrase cache seed byte length
+ *     +24  4 bytes  ref_audio_tokens offset
+ *     +28  4 bytes  ref_audio_tokens byte length
+ *     +32  4 bytes  ref_text offset
+ *     +36  4 bytes  ref_text byte length
+ *     +40  4 bytes  instruct offset
+ *     +44  4 bytes  instruct byte length
+ *     +48  4 bytes  metadata offset
+ *     +52  4 bytes  metadata byte length
+ *     +56  4 bytes  reserved (must be 0)
+ *     +60  4 bytes  reserved (must be 0)
+ *
+ *   `ref_audio_tokens` payload (v2):
+ *     +0   4 bytes  K            — codebook count (uint32, OmniVoice = 8)
+ *     +4   4 bytes  ref_T        — frames per codebook (uint32)
+ *     +8   ...      int32 LE codebook samples, row-major shape `[K, ref_T]`
+ *
+ *   `ref_text` payload (v2): raw UTF-8 bytes (no NUL terminator).
+ *   `instruct` payload (v2): raw UTF-8 bytes (closed VoiceDesign vocabulary).
+ *   `metadata` payload (v2): raw UTF-8 JSON bytes (codec sha256, corpus
+ *                            hash, etc.); the runtime never relies on
+ *                            metadata for correctness.
+ *
+ *   Phrase cache seed payload (v1 + v2, identical):
+ *     uint32 LE  N (phrase count)
+ *     for each phrase:
+ *       uint16 LE  text_byte_len
+ *       uint8[]    canonicalized text (UTF-8)
+ *       uint32 LE  sample_rate
+ *       uint32 LE  pcm_byte_len
+ *       uint8[]    PCM (Float32 LE samples)
+ *
+ * Per-section invariants:
+ *   - Section bounds may not overlap the header.
+ *   - Section bounds must fit within the file length.
+ *   - A `length=0` section is allowed (means "absent"); the corresponding
+ *     output field is an empty `Float32Array` / `Int32Array` / empty string.
+ *   - `embedding.length % 4 == 0` (Float32).
+ *   - `ref_audio_tokens.length` ≥ 8 (the two header words K, ref_T) and the
+ *     payload is `8 + K*ref_T*4` bytes.
+ */
+
+export const VOICE_PRESET_MAGIC = 0x315a4c45; // 'ELZ1'
+
+/** Header byte counts. */
+export const VOICE_PRESET_HEADER_BYTES_V1 = 24;
+export const VOICE_PRESET_HEADER_BYTES_V2 = 64;
+
+/** Supported format versions. v2 is the canonical write path. */
+export const VOICE_PRESET_VERSION_V1 = 1;
+export const VOICE_PRESET_VERSION_V2 = 2;
+export const VOICE_PRESET_VERSION_CURRENT = VOICE_PRESET_VERSION_V2;
+
+export interface VoicePresetSeedPhrase {
+	/** Canonicalized text (lowercase, single-spaced, trimmed). */
+	text: string;
+	sampleRate: number;
+	pcm: Float32Array;
+}
+
+/**
+ * OmniVoice reference-audio-tokens payload. `K` is the codebook count (=8 for
+ * OmniVoice / HiggsAudioV2) and `refT` is the number of frames per codebook.
+ * `tokens` is row-major: codebook `k`, frame `t` is at `tokens[k*refT + t]`.
+ * An empty payload (refT=0, K=0, tokens length 0) is valid and means "no
+ * reference audio bound to this preset" (instruct-only voice).
+ */
+export interface RefAudioTokens {
+	K: number;
+	refT: number;
+	tokens: Int32Array;
+}
+
+export interface VoicePresetFile {
+	version: number;
+	embedding: Float32Array;
+	phrases: ReadonlyArray<VoicePresetSeedPhrase>;
+	/** v2 only — empty for v1 files. */
+	refAudioTokens: RefAudioTokens;
+	/** v2 only — empty for v1 files. */
+	refText: string;
+	/** v2 only — empty for v1 files. */
+	instruct: string;
+	/** v2 only — parsed JSON object, empty `{}` for v1 files. */
+	metadata: Record<string, unknown>;
+}
+
+export class VoicePresetFormatError extends Error {
+	constructor(
+		message: string,
+		readonly code:
+			| "bad-magic"
+			| "bad-version"
+			| "truncated-header"
+			| "truncated-section"
+			| "bad-section-bounds"
+			| "bad-phrase-record"
+			| "bad-embedding-length"
+			| "bad-ref-tokens"
+			| "bad-metadata",
+	) {
+		super(message);
+		this.name = "VoicePresetFormatError";
+	}
+}
+
+interface SectionView {
+	offset: number;
+	length: number;
+}
+
+interface ParsedHeader {
+	version: number;
+	headerBytes: number;
+	embedding: SectionView;
+	phrases: SectionView;
+	refAudioTokens: SectionView;
+	refText: SectionView;
+	instruct: SectionView;
+	metadata: SectionView;
+}
+
+const EMPTY_SECTION: SectionView = Object.freeze({ offset: 0, length: 0 });
+
+function checkSectionBounds(
+	sec: SectionView,
+	fileLen: number,
+	headerBytes: number,
+): void {
+	if (sec.length === 0) return;
+	if (sec.offset < headerBytes) {
+		throw new VoicePresetFormatError(
+			`voice preset section overlaps header (offset=${sec.offset} < header=${headerBytes})`,
+			"bad-section-bounds",
+		);
+	}
+	if (sec.offset + sec.length > fileLen) {
+		throw new VoicePresetFormatError(
+			`voice preset section bounds exceed file length`,
+			"bad-section-bounds",
+		);
+	}
+}
+
+function readHeader(view: DataView): ParsedHeader {
+	if (view.byteLength < VOICE_PRESET_HEADER_BYTES_V1) {
+		throw new VoicePresetFormatError(
+			`voice preset file truncated: header needs ${VOICE_PRESET_HEADER_BYTES_V1} bytes, got ${view.byteLength}`,
+			"truncated-header",
+		);
+	}
+	const magic = view.getUint32(0, true);
+	if (magic !== VOICE_PRESET_MAGIC) {
+		throw new VoicePresetFormatError(
+			`voice preset bad magic: expected 0x${VOICE_PRESET_MAGIC.toString(16)}, got 0x${magic.toString(16)}`,
+			"bad-magic",
+		);
+	}
+	const version = view.getUint32(4, true);
+	if (
+		version !== VOICE_PRESET_VERSION_V1 &&
+		version !== VOICE_PRESET_VERSION_V2
+	) {
+		throw new VoicePresetFormatError(
+			`voice preset unsupported version: ${version} (this build supports 1 and 2)`,
+			"bad-version",
+		);
+	}
+	const headerBytes =
+		version === VOICE_PRESET_VERSION_V2
+			? VOICE_PRESET_HEADER_BYTES_V2
+			: VOICE_PRESET_HEADER_BYTES_V1;
+	if (view.byteLength < headerBytes) {
+		throw new VoicePresetFormatError(
+			`voice preset file truncated: v${version} header needs ${headerBytes} bytes, got ${view.byteLength}`,
+			"truncated-header",
+		);
+	}
+
+	const embedding: SectionView = {
+		offset: view.getUint32(8, true),
+		length: view.getUint32(12, true),
+	};
+	const phrases: SectionView = {
+		offset: view.getUint32(16, true),
+		length: view.getUint32(20, true),
+	};
+
+	let refAudioTokens = EMPTY_SECTION;
+	let refText = EMPTY_SECTION;
+	let instruct = EMPTY_SECTION;
+	let metadata = EMPTY_SECTION;
+	if (version === VOICE_PRESET_VERSION_V2) {
+		refAudioTokens = {
+			offset: view.getUint32(24, true),
+			length: view.getUint32(28, true),
+		};
+		refText = {
+			offset: view.getUint32(32, true),
+			length: view.getUint32(36, true),
+		};
+		instruct = {
+			offset: view.getUint32(40, true),
+			length: view.getUint32(44, true),
+		};
+		metadata = {
+			offset: view.getUint32(48, true),
+			length: view.getUint32(52, true),
+		};
+		// Reserved words must be zero — fail closed on accidental reuse.
+		const r0 = view.getUint32(56, true);
+		const r1 = view.getUint32(60, true);
+		if (r0 !== 0 || r1 !== 0) {
+			throw new VoicePresetFormatError(
+				`voice preset v2 reserved header words must be 0 (got ${r0}, ${r1})`,
+				"bad-section-bounds",
+			);
+		}
+	}
+
+	const fileLen = view.byteLength;
+	checkSectionBounds(embedding, fileLen, headerBytes);
+	checkSectionBounds(phrases, fileLen, headerBytes);
+	checkSectionBounds(refAudioTokens, fileLen, headerBytes);
+	checkSectionBounds(refText, fileLen, headerBytes);
+	checkSectionBounds(instruct, fileLen, headerBytes);
+	checkSectionBounds(metadata, fileLen, headerBytes);
+
+	return {
+		version,
+		headerBytes,
+		embedding,
+		phrases,
+		refAudioTokens,
+		refText,
+		instruct,
+		metadata,
+	};
+}
+
+function copyFloat32(
+	bytes: Uint8Array,
+	/** Offset relative to `bytes` (i.e. relative to bytes.byteOffset). */
+	relativeOffset: number,
+	byteLength: number,
+): Float32Array {
+	// The source byte offset is not guaranteed to be 4-aligned in the file
+	// buffer, so we copy raw bytes into a fresh ArrayBuffer first.
+	const aligned = new Uint8Array(byteLength);
+	aligned.set(bytes.subarray(relativeOffset, relativeOffset + byteLength));
+	return new Float32Array(aligned.buffer, 0, byteLength / 4);
+}
+
+function copyInt32(
+	bytes: Uint8Array,
+	relativeOffset: number,
+	byteLength: number,
+): Int32Array {
+	const aligned = new Uint8Array(byteLength);
+	aligned.set(bytes.subarray(relativeOffset, relativeOffset + byteLength));
+	return new Int32Array(aligned.buffer, 0, byteLength / 4);
+}
+
+function readEmbedding(bytes: Uint8Array, sec: SectionView): Float32Array {
+	if (sec.length === 0) return new Float32Array(0);
+	if (sec.length % 4 !== 0) {
+		throw new VoicePresetFormatError(
+			`voice preset embedding length ${sec.length} is not a multiple of 4`,
+			"bad-embedding-length",
+		);
+	}
+	return copyFloat32(bytes, sec.offset, sec.length);
+}
+
+function readRefAudioTokens(
+	bytes: Uint8Array,
+	sec: SectionView,
+): RefAudioTokens {
+	if (sec.length === 0) {
+		return { K: 0, refT: 0, tokens: new Int32Array(0) };
+	}
+	if (sec.length < 8) {
+		throw new VoicePresetFormatError(
+			`voice preset ref_audio_tokens section truncated (need ≥ 8 bytes, got ${sec.length})`,
+			"bad-ref-tokens",
+		);
+	}
+	const view = new DataView(
+		bytes.buffer,
+		bytes.byteOffset + sec.offset,
+		sec.length,
+	);
+	const K = view.getUint32(0, true);
+	const refT = view.getUint32(4, true);
+	const tokenBytes = sec.length - 8;
+	if (tokenBytes % 4 !== 0) {
+		throw new VoicePresetFormatError(
+			`voice preset ref_audio_tokens payload bytes ${tokenBytes} is not a multiple of 4`,
+			"bad-ref-tokens",
+		);
+	}
+	const expected = K * refT * 4;
+	if (tokenBytes !== expected) {
+		throw new VoicePresetFormatError(
+			`voice preset ref_audio_tokens shape mismatch: K=${K}, ref_T=${refT}, expected ${expected} bytes, got ${tokenBytes}`,
+			"bad-ref-tokens",
+		);
+	}
+	const tokens =
+		tokenBytes === 0
+			? new Int32Array(0)
+			: copyInt32(bytes, sec.offset + 8, tokenBytes);
+	return { K, refT, tokens };
+}
+
+function readUtf8(bytes: Uint8Array, sec: SectionView): string {
+	if (sec.length === 0) return "";
+	const slice = bytes.subarray(sec.offset, sec.offset + sec.length);
+	return new TextDecoder("utf-8", { fatal: true }).decode(slice);
+}
+
+function readMetadata(
+	bytes: Uint8Array,
+	sec: SectionView,
+): Record<string, unknown> {
+	if (sec.length === 0) return {};
+	const text = readUtf8(bytes, sec);
+	let parsed: unknown;
+	try {
+		parsed = JSON.parse(text);
+	} catch (err) {
+		throw new VoicePresetFormatError(
+			`voice preset metadata is not valid JSON: ${(err as Error).message}`,
+			"bad-metadata",
+		);
+	}
+	if (parsed === null || typeof parsed !== "object" || Array.isArray(parsed)) {
+		throw new VoicePresetFormatError(
+			`voice preset metadata must be a JSON object`,
+			"bad-metadata",
+		);
+	}
+	return parsed as Record<string, unknown>;
+}
+
+function readPhrases(
+	bytes: Uint8Array,
+	sec: SectionView,
+): VoicePresetSeedPhrase[] {
+	if (sec.length === 0) return [];
+	const view = new DataView(
+		bytes.buffer,
+		bytes.byteOffset + sec.offset,
+		sec.length,
+	);
+	const decoder = new TextDecoder("utf-8", { fatal: true });
+	let pos = 0;
+	if (sec.length < 4) {
+		throw new VoicePresetFormatError(
+			"voice preset phrase section truncated before count",
+			"truncated-section",
+		);
+	}
+	const count = view.getUint32(pos, true);
+	pos += 4;
+	const out: VoicePresetSeedPhrase[] = [];
+	for (let i = 0; i < count; i++) {
+		if (pos + 2 > sec.length) {
+			throw new VoicePresetFormatError(
+				`voice preset phrase #${i}: truncated before text length`,
+				"bad-phrase-record",
+			);
+		}
+		const textLen = view.getUint16(pos, true);
+		pos += 2;
+		if (pos + textLen > sec.length) {
+			throw new VoicePresetFormatError(
+				`voice preset phrase #${i}: text overruns section`,
+				"bad-phrase-record",
+			);
+		}
+		const textBytes = new Uint8Array(
+			bytes.buffer,
+			bytes.byteOffset + sec.offset + pos,
+			textLen,
+		);
+		const text = decoder.decode(textBytes);
+		pos += textLen;
+		if (pos + 8 > sec.length) {
+			throw new VoicePresetFormatError(
+				`voice preset phrase #${i}: truncated before sample_rate/pcm_len`,
+				"bad-phrase-record",
+			);
+		}
+		const sampleRate = view.getUint32(pos, true);
+		pos += 4;
+		const pcmByteLen = view.getUint32(pos, true);
+		pos += 4;
+		if (pcmByteLen % 4 !== 0) {
+			throw new VoicePresetFormatError(
+				`voice preset phrase #${i}: pcm byte length ${pcmByteLen} is not a multiple of 4`,
+				"bad-phrase-record",
+			);
+		}
+		if (pos + pcmByteLen > sec.length) {
+			throw new VoicePresetFormatError(
+				`voice preset phrase #${i}: pcm overruns section`,
+				"bad-phrase-record",
+			);
+		}
+		const pcm = copyFloat32(bytes, sec.offset + pos, pcmByteLen);
+		pos += pcmByteLen;
+		out.push({ text, sampleRate, pcm });
+	}
+	return out;
+}
+
+/**
+ * Parse a voice-preset binary blob. Throws `VoicePresetFormatError` on any
+ * malformed input — this is the single defensive boundary for the format.
+ * Supports both v1 and v2 files. For v1 files the v2-only fields are
+ * returned as their empty equivalents.
+ */
+export function readVoicePresetFile(bytes: Uint8Array): VoicePresetFile {
+	const view = new DataView(bytes.buffer, bytes.byteOffset, bytes.byteLength);
+	const header = readHeader(view);
+	return {
+		version: header.version,
+		embedding: readEmbedding(bytes, header.embedding),
+		phrases: readPhrases(bytes, header.phrases),
+		refAudioTokens: readRefAudioTokens(bytes, header.refAudioTokens),
+		refText: readUtf8(bytes, header.refText),
+		instruct: readUtf8(bytes, header.instruct),
+		metadata: readMetadata(bytes, header.metadata),
+	};
+}
+
+/**
+ * Serialize a voice preset to the v1 binary format. The output is a fresh
+ * `Uint8Array` ready to be written to disk.
+ *
+ * Use this only when the caller deliberately wants the legacy v1 shape (e.g.
+ * the existing Kokoro-style placeholder builder). New code should call
+ * `writeVoicePresetFileV2`.
+ */
+export function writeVoicePresetFile(file: {
+	embedding: Float32Array;
+	phrases: ReadonlyArray<VoicePresetSeedPhrase>;
+}): Uint8Array {
+	const encoder = new TextEncoder();
+	const encodedTexts = file.phrases.map((p) => encoder.encode(p.text));
+
+	const embBytes = file.embedding.byteLength;
+	let phrBytes = 4; // count
+	for (let i = 0; i < file.phrases.length; i++) {
+		const t = encodedTexts[i];
+		if (t.byteLength > 0xffff) {
+			throw new VoicePresetFormatError(
+				`phrase #${i} text too long (${t.byteLength} bytes, max 65535)`,
+				"bad-phrase-record",
+			);
+		}
+		phrBytes += 2 + t.byteLength + 4 + 4 + file.phrases[i].pcm.byteLength;
+	}
+
+	const embOff = VOICE_PRESET_HEADER_BYTES_V1;
+	const phrOff = embOff + embBytes;
+	const total = phrOff + phrBytes;
+
+	const out = new Uint8Array(total);
+	const view = new DataView(out.buffer);
+	view.setUint32(0, VOICE_PRESET_MAGIC, true);
+	view.setUint32(4, VOICE_PRESET_VERSION_V1, true);
+	view.setUint32(8, embOff, true);
+	view.setUint32(12, embBytes, true);
+	view.setUint32(16, phrOff, true);
+	view.setUint32(20, phrBytes, true);
+
+	// Embedding
+	out.set(
+		new Uint8Array(
+			file.embedding.buffer,
+			file.embedding.byteOffset,
+			file.embedding.byteLength,
+		),
+		embOff,
+	);
+
+	// Phrases
+	writePhraseSection(out, view, phrOff, file.phrases, encodedTexts);
+
+	return out;
+}
+
+function writePhraseSection(
+	out: Uint8Array,
+	view: DataView,
+	startOff: number,
+	phrases: ReadonlyArray<VoicePresetSeedPhrase>,
+	encodedTexts: Uint8Array[],
+): void {
+	let pos = startOff;
+	view.setUint32(pos, phrases.length, true);
+	pos += 4;
+	for (let i = 0; i < phrases.length; i++) {
+		const t = encodedTexts[i];
+		const phrase = phrases[i];
+		view.setUint16(pos, t.byteLength, true);
+		pos += 2;
+		out.set(t, pos);
+		pos += t.byteLength;
+		view.setUint32(pos, phrase.sampleRate, true);
+		pos += 4;
+		view.setUint32(pos, phrase.pcm.byteLength, true);
+		pos += 4;
+		out.set(
+			new Uint8Array(
+				phrase.pcm.buffer,
+				phrase.pcm.byteOffset,
+				phrase.pcm.byteLength,
+			),
+			pos,
+		);
+		pos += phrase.pcm.byteLength;
+	}
+}
+
+/**
+ * Write a voice preset in the v2 (additive) layout. Used by the OmniVoice
+ * freeze pipeline (`freeze-voice.mjs`) and other producers that need to
+ * persist `refAudioTokens` / `refText` / `instruct` alongside the v1
+ * embedding + phrase-seed sections.
+ *
+ * Any field that the caller doesn't need to persist can be omitted (or
+ * passed empty). The on-disk section is then written as length=0 and is
+ * read back as the empty equivalent.
+ */
+export function writeVoicePresetFileV2(file: {
+	embedding?: Float32Array;
+	phrases?: ReadonlyArray<VoicePresetSeedPhrase>;
+	refAudioTokens?: RefAudioTokens;
+	refText?: string;
+	instruct?: string;
+	metadata?: Record<string, unknown>;
+}): Uint8Array {
+	const embedding = file.embedding ?? new Float32Array(0);
+	const phrases = file.phrases ?? [];
+	const refAudioTokens = file.refAudioTokens ?? {
+		K: 0,
+		refT: 0,
+		tokens: new Int32Array(0),
+	};
+	const refText = file.refText ?? "";
+	const instruct = file.instruct ?? "";
+	const metadata = file.metadata ?? {};
+
+	if (refAudioTokens.K * refAudioTokens.refT !== refAudioTokens.tokens.length) {
+		throw new VoicePresetFormatError(
+			`ref_audio_tokens shape mismatch: K=${refAudioTokens.K}, ref_T=${refAudioTokens.refT}, but tokens.length=${refAudioTokens.tokens.length}`,
+			"bad-ref-tokens",
+		);
+	}
+
+	const encoder = new TextEncoder();
+	const encodedTexts = phrases.map((p) => encoder.encode(p.text));
+	const encodedRefText = encoder.encode(refText);
+	const encodedInstruct = encoder.encode(instruct);
+	const encodedMetadata =
+		Object.keys(metadata).length === 0
+			? new Uint8Array(0)
+			: encoder.encode(JSON.stringify(metadata));
+
+	// Compute payload sizes up-front so we can lay out section offsets.
+	const embBytes = embedding.byteLength;
+	let phrBytes = phrases.length === 0 && encodedTexts.length === 0 ? 0 : 4;
+	if (phrBytes > 0) {
+		for (let i = 0; i < phrases.length; i++) {
+			const t = encodedTexts[i];
+			if (t.byteLength > 0xffff) {
+				throw new VoicePresetFormatError(
+					`phrase #${i} text too long (${t.byteLength} bytes, max 65535)`,
+					"bad-phrase-record",
+				);
+			}
+			phrBytes += 2 + t.byteLength + 4 + 4 + phrases[i].pcm.byteLength;
+		}
+	}
+	const refTokensBytes =
+		refAudioTokens.tokens.length === 0 && refAudioTokens.K === 0
+			? 0
+			: 8 + refAudioTokens.tokens.byteLength;
+
+	// Lay out sections in declared order. Empty sections claim no space and
+	// are recorded as (offset=0, length=0).
+	let cursor = VOICE_PRESET_HEADER_BYTES_V2;
+	const embOff = embBytes > 0 ? cursor : 0;
+	cursor += embBytes;
+	const phrOff = phrBytes > 0 ? cursor : 0;
+	cursor += phrBytes;
+	const refTokensOff = refTokensBytes > 0 ? cursor : 0;
+	cursor += refTokensBytes;
+	const refTextOff = encodedRefText.byteLength > 0 ? cursor : 0;
+	cursor += encodedRefText.byteLength;
+	const instructOff = encodedInstruct.byteLength > 0 ? cursor : 0;
+	cursor += encodedInstruct.byteLength;
+	const metadataOff = encodedMetadata.byteLength > 0 ? cursor : 0;
+	cursor += encodedMetadata.byteLength;
+
+	const total = cursor;
+	const out = new Uint8Array(total);
+	const view = new DataView(out.buffer);
+
+	view.setUint32(0, VOICE_PRESET_MAGIC, true);
+	view.setUint32(4, VOICE_PRESET_VERSION_V2, true);
+	view.setUint32(8, embOff, true);
+	view.setUint32(12, embBytes, true);
+	view.setUint32(16, phrOff, true);
+	view.setUint32(20, phrBytes, true);
+	view.setUint32(24, refTokensOff, true);
+	view.setUint32(28, refTokensBytes, true);
+	view.setUint32(32, refTextOff, true);
+	view.setUint32(36, encodedRefText.byteLength, true);
+	view.setUint32(40, instructOff, true);
+	view.setUint32(44, encodedInstruct.byteLength, true);
+	view.setUint32(48, metadataOff, true);
+	view.setUint32(52, encodedMetadata.byteLength, true);
+	view.setUint32(56, 0, true);
+	view.setUint32(60, 0, true);
+
+	if (embBytes > 0) {
+		out.set(
+			new Uint8Array(embedding.buffer, embedding.byteOffset, embBytes),
+			embOff,
+		);
+	}
+	if (phrBytes > 0) {
+		writePhraseSection(out, view, phrOff, phrases, encodedTexts);
+	}
+	if (refTokensBytes > 0) {
+		view.setUint32(refTokensOff, refAudioTokens.K, true);
+		view.setUint32(refTokensOff + 4, refAudioTokens.refT, true);
+		if (refAudioTokens.tokens.byteLength > 0) {
+			out.set(
+				new Uint8Array(
+					refAudioTokens.tokens.buffer,
+					refAudioTokens.tokens.byteOffset,
+					refAudioTokens.tokens.byteLength,
+				),
+				refTokensOff + 8,
+			);
+		}
+	}
+	if (encodedRefText.byteLength > 0) {
+		out.set(encodedRefText, refTextOff);
+	}
+	if (encodedInstruct.byteLength > 0) {
+		out.set(encodedInstruct, instructOff);
+	}
+	if (encodedMetadata.byteLength > 0) {
+		out.set(encodedMetadata, metadataOff);
+	}
+
+	return out;
+}