carta1 1.0.12 → 1.0.14

package/codec/coding/bitallocation.js

@@ -17,7 +17,11 @@ import {
   NUM_BFUS,
   SAMPLE_RATE,
   BFU_FREQUENCIES,
+  BFU_AMOUNTS_COUNT,
   PSYMODEL_MIN_POWER_DB,
+  PSYMODEL_CB_FREQ_INDICES,
+  INTERPOLATION_COMPENSATION_FACTOR,
+  PSYMODEL_FFT_SIZE,
 } from '../core/constants.js'
 
 /**
@@ -32,10 +36,10 @@ export function allocateBits(psychoResults, bfuData, bfuSizes, maxBfuCount) {
   // Calculate signal energy for each BFU
   const signalEnergyDb = calculateBfuEnergy(bfuData, bfuSizes, maxBfuCount)
 
-  // Map psychoacoustic masking threshold to BFUs
-  const maskingThreshold =
-    psychoResults.maskingThreshold || psychoResults.globalThreshold
-  const maskThresholdDb = mapThresholdToBfus(maskingThreshold)
+  // Interpolate critical band thresholds to BFUs
+  const maskThresholdDb = interpolateCriticalBandsToBfus(
+    psychoResults.criticalBandThresholds
+  )
 
   // Calculate Signal-to-Mask Ratios
   const smrValues = calculateSMR(signalEnergyDb, maskThresholdDb, maxBfuCount)
@@ -90,7 +94,7 @@ function calculateBfuEnergy(bfuData, bfuSizes, bfuCount) {
       energy += data[j] * data[j]
     }
 
-    // Convert to dB (10 * log10(energy/size))
+    // Convert to dB correctly using log10
     energyPerBfu[i] =
       energy > 0 ? 10 * Math.log10(energy / size) : PSYMODEL_MIN_POWER_DB
   }
@@ -99,25 +103,51 @@ function calculateBfuEnergy(bfuData, bfuSizes, bfuCount) {
 }
 
 /**
- * Map frequency-domain masking threshold to BFU indices
- * @param {Float32Array} globalThreshold - Frequency-domain masking threshold
+ * Interpolate critical band thresholds directly to BFU frequencies
+ * @param {Float32Array} criticalBandThresholds - Critical band thresholds
  * @returns {Float32Array} Masking threshold per BFU
  */
-function mapThresholdToBfus(globalThreshold) {
+function interpolateCriticalBandsToBfus(criticalBandThresholds) {
   const maskThresholdPerBfu = new Float32Array(NUM_BFUS)
-  maskThresholdPerBfu.fill(PSYMODEL_MIN_POWER_DB)
-
-  // Calculate frequency resolution
-  const fftSize = (globalThreshold.length - 1) * 2
-  const freqPerBin = SAMPLE_RATE / fftSize
+  const freqPerBin = SAMPLE_RATE / PSYMODEL_FFT_SIZE
 
-  // Map each BFU's center frequency to nearest FFT bin
   for (let i = 0; i < NUM_BFUS; i++) {
     const bfuFreq = BFU_FREQUENCIES[i]
     const fftBinIndex = Math.round(bfuFreq / freqPerBin)
 
-    if (fftBinIndex >= 0 && fftBinIndex < globalThreshold.length) {
-      maskThresholdPerBfu[i] = globalThreshold[fftBinIndex]
+    // Find which critical bands surround this FFT bin
+    let bandIndex = 0
+    while (
+      bandIndex < PSYMODEL_CB_FREQ_INDICES.length - 1 &&
+      PSYMODEL_CB_FREQ_INDICES[bandIndex + 1] < fftBinIndex
+    ) {
+      bandIndex++
+    }
+
+    if (fftBinIndex <= PSYMODEL_CB_FREQ_INDICES[0]) {
+      maskThresholdPerBfu[i] = criticalBandThresholds[0]
+    } else if (
+      fftBinIndex >=
+      PSYMODEL_CB_FREQ_INDICES[PSYMODEL_CB_FREQ_INDICES.length - 1]
+    ) {
+      maskThresholdPerBfu[i] =
+        criticalBandThresholds[criticalBandThresholds.length - 1]
+    } else {
+      const x0 = PSYMODEL_CB_FREQ_INDICES[bandIndex]
+      const x1 = PSYMODEL_CB_FREQ_INDICES[bandIndex + 1]
+      const y0 = criticalBandThresholds[bandIndex]
+      const y1 = criticalBandThresholds[bandIndex + 1]
+      const bandWidth = x1 - x0
+
+      if (bandWidth > 0) {
+        const t = (fftBinIndex - x0) / bandWidth
+        const interpolated = y0 + (y1 - y0) * t
+        const compensation =
+          Math.abs(y1 - y0) * INTERPOLATION_COMPENSATION_FACTOR
+        maskThresholdPerBfu[i] = interpolated + compensation
+      } else {
+        maskThresholdPerBfu[i] = y0
+      }
     }
   }
 
@@ -135,7 +165,6 @@ function calculateSMR(signalEnergyDb, maskThresholdDb, maxBfuCount) {
   const smrValues = new Float32Array(maxBfuCount)
 
   for (let bfu = 0; bfu < maxBfuCount; bfu++) {
-    // BFUs with no signal have undefined SMR
     smrValues[bfu] =
       signalEnergyDb[bfu] <= PSYMODEL_MIN_POWER_DB
         ? -Infinity
@@ -157,111 +186,86 @@ function selectOptimalBfuCount(smrValues, signalEnergyDb, maxBfuCount) {
   const hasAudibleContent = smrValues.some((smr) => smr > 0)
 
   if (hasAudibleContent) {
-    return selectBfuCountForAudibleContent(smrValues, maxBfuCount)
-  } else {
-    return selectBfuCountForInaudibleContent(
-      smrValues,
-      signalEnergyDb,
-      maxBfuCount
-    )
-  }
-}
-
-/**
- * Select BFU count when content is above masking threshold
- * Uses diminishing returns approach - find where adding more BFUs
- * provides minimal benefit
- */
-function selectBfuCountForAudibleContent(smrValues, maxBfuCount) {
-  // Create sorted list of BFUs by their contribution (positive SMR only)
-  const contributions = []
-  for (let i = 0; i < maxBfuCount; i++) {
-    if (smrValues[i] > 0) {
-      contributions.push({ index: i, smr: smrValues[i] })
+    // Create sorted list of BFUs by their contribution (positive SMR only)
+    const contributions = []
+    for (let i = 0; i < maxBfuCount; i++) {
+      if (smrValues[i] > 0) {
+        contributions.push({ index: i, smr: smrValues[i] })
+      }
     }
-  }
-  contributions.sort((a, b) => b.smr - a.smr)
-
-  if (contributions.length === 0) {
-    return BFU_AMOUNTS[BFU_AMOUNTS.length - 1]
-  }
-
-  // Find the natural cutoff using gradient of cumulative SMR
-  const cumulativeSMR = []
-  let sum = 0
-  for (let i = 0; i < contributions.length; i++) {
-    sum += contributions[i].smr
-    cumulativeSMR.push(sum)
-  }
+    contributions.sort((a, b) => b.smr - a.smr)
 
-  // Find where the gradient becomes shallow (diminishing returns)
-  for (const candidateCount of BFU_AMOUNTS) {
-    if (candidateCount > maxBfuCount) continue
-
-    // Check if all significant contributions fit within this count
-    const requiredBfus = contributions
-      .slice(0, Math.min(contributions.length, candidateCount))
-      .map((c) => c.index)
+    if (contributions.length === 0) {
+      return BFU_AMOUNTS[BFU_AMOUNTS.length - 1]
+    }
 
-    const maxRequiredIndex = Math.max(...requiredBfus, 0)
+    // Find the natural cutoff using gradient of cumulative SMR
+    const cumulativeSMR = []
+    let sum = 0
+    for (let i = 0; i < contributions.length; i++) {
+      sum += contributions[i].smr
+      cumulativeSMR.push(sum)
+    }
 
-    if (maxRequiredIndex < candidateCount) {
-      // Check if we've captured the bulk of the contribution
-      const capturedIndex = Math.min(
-        candidateCount - 1,
-        contributions.length - 1
-      )
-      if (capturedIndex >= contributions.length - 1) {
-        return candidateCount
-      }
+    // Find where the gradient becomes shallow (diminishing returns)
+    for (const candidateCount of BFU_AMOUNTS) {
+      if (candidateCount > maxBfuCount) continue
 
-      // Use gradient to detect diminishing returns
-      const currentValue = cumulativeSMR[capturedIndex]
-      const totalValue = cumulativeSMR[cumulativeSMR.length - 1]
-      const remainingIndices = contributions.length - capturedIndex - 1
+      // Check if all significant contributions fit within this count
+      const requiredBfus = contributions
+        .slice(0, Math.min(contributions.length, candidateCount))
+        .map((c) => c.index)
 
-      if (remainingIndices > 0) {
-        const averageRemaining = (totalValue - currentValue) / remainingIndices
-        const averageCaptured = currentValue / (capturedIndex + 1)
+      const maxRequiredIndex = Math.max(...requiredBfus, 0)
 
-        // If remaining BFUs contribute much less on average, we've found our cutoff
-        if (averageRemaining < averageCaptured * 0.1) {
+      if (maxRequiredIndex < candidateCount) {
+        // Check if we've captured the bulk of the contribution
+        const capturedIndex = Math.min(
+          candidateCount - 1,
+          contributions.length - 1
+        )
+        if (capturedIndex >= contributions.length - 1) {
           return candidateCount
         }
+
+        // Use gradient to detect diminishing returns
+        const remainingIndices = contributions.length - capturedIndex - 1
+        if (remainingIndices > 0) {
+          const avgRemaining =
+            (cumulativeSMR[cumulativeSMR.length - 1] -
+              cumulativeSMR[capturedIndex]) /
+            remainingIndices
+          const avgCaptured = cumulativeSMR[capturedIndex] / (capturedIndex + 1)
+
+          // If remaining BFUs contribute much less on average, we've found our cutoff
+          if (avgRemaining < avgCaptured * 0.1) {
+            return candidateCount
+          }
+        }
       }
     }
-  }
-
-  return BFU_AMOUNTS[BFU_AMOUNTS.length - 1]
-}
-
-/**
- * Select BFU count when all content is below masking threshold
- * Simply finds the smallest count that includes all BFUs with signal
- */
-function selectBfuCountForInaudibleContent(
-  smrValues,
-  signalEnergyDb,
-  maxBfuCount
-) {
-  // Find the highest BFU index that contains actual signal
-  let highestSignalBfu = -1
-
-  for (let i = maxBfuCount - 1; i >= 0; i--) {
-    if (signalEnergyDb[i] > PSYMODEL_MIN_POWER_DB && smrValues[i] > -Infinity) {
-      highestSignalBfu = i
-      break
+  } else {
+    // Find the highest BFU index that contains actual signal
+    let highestSignalBfu = -1
+    for (let i = maxBfuCount - 1; i >= 0; i--) {
+      if (
+        signalEnergyDb[i] > PSYMODEL_MIN_POWER_DB &&
+        smrValues[i] > -Infinity
+      ) {
+        highestSignalBfu = i
+        break
+      }
     }
-  }
 
-  // Find smallest BFU count that includes all signal
-  for (const candidateCount of BFU_AMOUNTS) {
-    if (candidateCount > highestSignalBfu) {
-      return candidateCount
+    // Find smallest BFU count that includes all signal
+    for (const candidateCount of BFU_AMOUNTS) {
+      if (candidateCount > highestSignalBfu) {
+        return candidateCount
+      }
     }
   }
 
-  return BFU_AMOUNTS[BFU_AMOUNTS.length - 1]
+  return BFU_AMOUNTS[BFU_AMOUNTS_COUNT - 1]
 }
 
 /**
@@ -278,36 +282,44 @@ function performBitAllocation(
   bfuSizes,
   availableBits
 ) {
-  const allocation = new Int32Array(actualBfuCount).fill(0)
+  const allocation = new Int32Array(actualBfuCount)
+  const heap = []
+
+  // Initialize heap with BFUs that have signal
+  for (let i = 0; i < actualBfuCount; i++) {
+    if (smrValues[i] > -Infinity) {
+      heap.push({ bfu: i, priority: smrValues[i] })
+    }
+  }
 
-  // Initialize priority queue with BFUs that have signal
-  const priorityQueue = createPriorityQueue(smrValues, actualBfuCount)
+  // Build max heap
+  for (let i = Math.floor(heap.length / 2) - 1; i >= 0; i--) {
+    heapifyDown(heap, i)
+  }
 
-  // Greedy allocation: always upgrade BFU with highest priority
-  while (!priorityQueue.isEmpty()) {
-    const bestBfu = priorityQueue.peek()
+  // Greedy allocation
+  while (heap.length > 0 && availableBits > 0) {
+    const bestBfu = heap[0]
     const currentWl = allocation[bestBfu.bfu]
 
-    // Check if BFU is already at maximum word length
     if (currentWl >= MAX_WORD_LENGTH_INDEX) {
-      priorityQueue.pop()
+      heap[0] = heap[heap.length - 1]
+      heap.pop()
+      if (heap.length > 0) heapifyDown(heap, 0)
       continue
     }
 
-    // Calculate cost to upgrade this BFU
     const upgradeCost = calculateUpgradeCost(bestBfu.bfu, currentWl, bfuSizes)
 
     if (upgradeCost <= availableBits) {
-      // Perform upgrade
       availableBits -= upgradeCost
       allocation[bestBfu.bfu]++
-
-      // Update priority based on distortion reduction
       bestBfu.priority += DISTORTION_DELTA_DB[currentWl]
-      priorityQueue.updateTop()
+      heapifyDown(heap, 0)
     } else {
-      // Not enough bits for this BFU
-      priorityQueue.pop()
+      heap[0] = heap[heap.length - 1]
+      heap.pop()
+      if (heap.length > 0) heapifyDown(heap, 0)
     }
   }
 
@@ -316,6 +328,10 @@ function performBitAllocation(
 
 /**
  * Calculate bit cost to upgrade a BFU's word length
+ * @param {number} bfuIndex - Index of the BFU to upgrade
+ * @param {number} currentWl - Current word length of the BFU
+ * @param {number[]} bfuSizes - Array of BFU sizes (number of coefficients per BFU)
+ * @returns {number} Additional bits required to upgrade word length by 1
  */
 function calculateUpgradeCost(bfuIndex, currentWl, bfuSizes) {
   const currentBits = WORD_LENGTH_BITS[currentWl] * bfuSizes[bfuIndex]
@@ -323,52 +339,21 @@ function calculateUpgradeCost(bfuIndex, currentWl, bfuSizes) {
   return nextBits - currentBits
 }
 
-/**
- * Create a priority queue for BFU allocation
- * @param {Float32Array} smrValues - Initial priorities (SMR values)
- * @param {number} bfuCount - Number of BFUs
- * @returns {Object} Priority queue interface
- */
-function createPriorityQueue(smrValues, bfuCount) {
-  const heap = []
-
-  // Initialize with BFUs that have signal
-  for (let i = 0; i < bfuCount; i++) {
-    if (smrValues[i] > -Infinity) {
-      heap.push({ bfu: i, priority: smrValues[i] })
-    }
-  }
-
-  // Build max heap
-  for (let i = Math.floor(heap.length / 2) - 1; i >= 0; i--) {
-    heapifyDown(heap, i)
-  }
-
-  return {
-    isEmpty: () => heap.length === 0,
-    peek: () => heap[0],
-    pop: () => popMax(heap),
-    updateTop: () => heapifyDown(heap, 0),
-  }
-}
-
 /**
  * Maintain max heap property by moving element down
  * @param {Array} heap - Heap array
  * @param {number} startIdx - Starting index
  */
 function heapifyDown(heap, startIdx) {
-  let idx = startIdx
-  const element = heap[idx]
   const length = heap.length
+  const element = heap[startIdx]
 
   while (true) {
-    const leftChild = 2 * idx + 1
-    const rightChild = 2 * idx + 2
-    let largest = idx
+    const leftChild = 2 * startIdx + 1
+    const rightChild = 2 * startIdx + 2
+    let largest = startIdx
     let largestPriority = element.priority
 
-    // Find largest among parent and children
     if (leftChild < length && heap[leftChild].priority > largestPriority) {
       largest = leftChild
       largestPriority = heap[leftChild].priority
@@ -378,30 +363,11 @@ function heapifyDown(heap, startIdx) {
       largest = rightChild
     }
 
-    // If parent is largest, we're done
-    if (largest === idx) break
-
-    // Swap and continue
-    heap[idx] = heap[largest]
-    idx = largest
-  }
-
-  heap[idx] = element
-}
-
-/**
- * Remove and return the maximum element from heap
- * @param {Array} heap - Heap array
- * @returns {Object} Maximum element
- */
-function popMax(heap) {
-  const max = heap[0]
-  const last = heap.pop()
+    if (largest === startIdx) break
 
-  if (heap.length > 0) {
-    heap[0] = last
-    heapifyDown(heap, 0)
+    heap[startIdx] = heap[largest]
+    startIdx = largest
   }
 
-  return max
+  heap[startIdx] = element
 }

package/codec/core/buffers.js

@@ -7,7 +7,6 @@ import {
   MAX_BFU_SIZE,
   QMF_DELAY,
   QMF_HIGH_BAND_DELAY,
-  PSYMODEL_FFT_SIZE,
 } from './constants.js'
 
 export class BufferPool {
@@ -84,12 +83,6 @@ export class BufferPool {
       256: new Float32Array(256),
     }
 
-    // Psychoacoustic model buffers
-    this.psychoAnalysis = {
-      flags: new Uint8Array(PSYMODEL_FFT_SIZE / 2),
-      threshold: new Float32Array(PSYMODEL_FFT_SIZE / 2 + 1),
-    }
-
     // Bit allocation buffers
     this.bfuData = new Array(NUM_BFUS)
     for (let i = 0; i < NUM_BFUS; i++) {

package/codec/core/constants.js

@@ -33,6 +33,7 @@ export const SPECS_PER_BFU = new Int32Array([
   20, 20, 20, 20, 20,
 ])
 
+export const BFU_AMOUNTS_COUNT = 8
 export const BFU_AMOUNTS = new Int32Array([20, 28, 32, 36, 40, 44, 48, 52])
 export const BFU_BAND_BOUNDARIES = new Int32Array([20, 36, 52])
 
@@ -116,6 +117,9 @@ export const QMF_ODD = (() => {
   return odd
 })()
 
+// Bit allocation
+export const INTERPOLATION_COMPENSATION_FACTOR = 0.25
+
 // Quantization
 export const MAX_WORD_LENGTH_INDEX = 15
 export const WORD_LENGTH_BITS = new Int32Array([
@@ -188,6 +192,53 @@ export const SPREADING_MATRIX = (() => {
 export const PSYMODEL_MIN_POWER_DB = -200
 export const PSYMODEL_FFT_SIZE = 2048
 
+// Pre-computes lookup tables for resampling the MDCT power spectrum into a PSD.
+// For each PSD bin, these tables provide the two source MDCT indices and the
+// linear interpolation weight needed to calculate the final power value.
+export const [
+  PSYMODEL_PSD_SOURCE_IDX0,
+  PSYMODEL_PSD_SOURCE_IDX1,
+  PSYMODEL_PSD_INTERP_WEIGHT,
+] = (() => {
+  const half = PSYMODEL_FFT_SIZE >>> 1
+  const sizeMinus1 = PSYMODEL_FFT_SIZE - 1
+  const halfDiv2 = half / 2
+  const scale = sizeMinus1 / half
+
+  const idx0 = new Uint16Array(half + 1)
+  const idx1 = new Uint16Array(half + 1)
+  const w1 = new Float32Array(half + 1)
+
+  for (let i = 0; i <= half; i++) {
+    const src = i * scale
+    const base = src | 0
+
+    const j0 = base
+    idx0[i] =
+      j0 < half
+        ? j0 < halfDiv2
+          ? j0
+          : half - 1 - (j0 - halfDiv2)
+        : sizeMinus1 - (j0 - half)
+
+    const j1 = base + 1
+    if (j1 >= PSYMODEL_FFT_SIZE) {
+      idx1[i] = idx0[i]
+    } else {
+      idx1[i] =
+        j1 < half
+          ? j1 < halfDiv2
+            ? j1
+            : half - 1 - (j1 - halfDiv2)
+          : sizeMinus1 - (j1 - half)
+    }
+
+    w1[i] = src - base
+  }
+
+  return [idx0, idx1, w1]
+})()
+
 // FFT sizes for transient detection
 export const FFT_SIZE_LOW = 256
 export const FFT_SIZE_MID = 256

package/codec/pipeline/encoder.js

@@ -31,7 +31,6 @@ import {
   FFT_SIZE_LOW,
   FFT_SIZE_MID,
   FFT_SIZE_HIGH,
-  PSYMODEL_FFT_SIZE,
   MDCT_SIZE_SHORT,
   MDCT_SIZE_MID,
   MDCT_SIZE_LONG,
@@ -300,12 +299,7 @@ export function quantizationStage(context) {
       bufferPool.bfuData
     )
 
-    const psychoResults = psychoAnalysis(
-      coefficients,
-      PSYMODEL_FFT_SIZE,
-      bufferPool.psychoAnalysis,
-      options.normalizationDb
-    )
+    const psychoResults = psychoAnalysis(coefficients, options.normalizationDb)
     const { bfuCount: selectedBfuCount, allocation } = allocateBits(
       psychoResults,
       bfuData,