speechflow 2.2.1 → 2.3.0

package/speechflow-cli/src/speechflow-node-a2a-rnnoise-wt.ts

@@ -1,6 +1,6 @@
 /*
 **  SpeechFlow - Speech Processing Flow Graph
-**  Copyright (c) 2024-2025 Dr. Ralf S. Engelschall <rse@engelschall.com>
+**  Copyright (c) 2024-2026 Dr. Ralf S. Engelschall <rse@engelschall.com>
 **  Licensed under GPL 3.0 <https://spdx.org/licenses/GPL-3.0-only>
 */
 
@@ -46,7 +46,7 @@ parentPort!.on("message", (msg) => {
         /*  convert back Float32Array to Int16Array  */
         const i16 = new Int16Array(data.length)
         for (let i = 0; i < data.length; i++)
-            i16[i] = Math.round(f32a[i])
+            i16[i] = Math.max(-32768, Math.min(32767, Math.round(f32a[i])))
 
         /*  send processed frame back to parent  */
         parentPort!.postMessage({ type: "process-done", id, data: i16 }, [ i16.buffer ])

package/speechflow-cli/src/speechflow-node-a2a-rnnoise.ts

@@ -1,6 +1,6 @@
 /*
 **  SpeechFlow - Speech Processing Flow Graph
-**  Copyright (c) 2024-2025 Dr. Ralf S. Engelschall <rse@engelschall.com>
+**  Copyright (c) 2024-2026 Dr. Ralf S. Engelschall <rse@engelschall.com>
 **  Licensed under GPL 3.0 <https://spdx.org/licenses/GPL-3.0-only>
 */
 
@@ -71,14 +71,27 @@ export default class SpeechFlowNodeA2ARNNoise extends SpeechFlowNode {
             })
         })
 
+        /*  track pending promises  */
+        const pending = new Map<string, {
+            resolve: (arr: Int16Array<ArrayBuffer>) => void,
+            reject:  (err: Error)                   => void
+        }>()
+
+        /*  reject all pending promises on worker exit  */
+        this.worker.on("exit", () => {
+            const err = new Error("worker terminated")
+            for (const cb of pending.values())
+                cb.reject(err)
+            pending.clear()
+        })
+
         /*  receive message from worker  */
-        const pending = new Map<string, (arr: Int16Array<ArrayBuffer>) => void>()
         this.worker.on("message", (msg: any) => {
             if (typeof msg === "object" && msg !== null && msg.type === "process-done") {
                 const cb = pending.get(msg.id)
                 pending.delete(msg.id)
                 if (cb)
-                    cb(msg.data)
+                    cb.resolve(msg.data)
                 else
                     this.log("warning", `RNNoise worker thread sent back unexpected id: ${msg.id}`)
             }
@@ -92,8 +105,8 @@ export default class SpeechFlowNodeA2ARNNoise extends SpeechFlowNode {
             if (this.closing)
                 return segment
             const id = `${seq++}`
-            return new Promise<Int16Array<ArrayBuffer>>((resolve) => {
-                pending.set(id, (segment) => { resolve(segment) })
+            return new Promise<Int16Array<ArrayBuffer>>((resolve, reject) => {
+                pending.set(id, { resolve, reject })
                 this.worker!.postMessage({ type: "process", id, data: segment }, [ segment.buffer ])
             })
         }
@@ -113,20 +126,19 @@ export default class SpeechFlowNodeA2ARNNoise extends SpeechFlowNode {
                     callback(new Error("invalid chunk payload type"))
                 else {
                     /*  convert Buffer into Int16Array  */
-                    const payload = util.convertBufToI16(chunk.payload)
+                    const payload = util.convertBufToI16(chunk.payload, self.config.audioLittleEndian)
 
                     /*  process Int16Array in necessary segments  */
                     util.processInt16ArrayInSegments(payload, self.sampleSize, (segment) =>
                         workerProcessSegment(segment)
                     ).then((payload: Int16Array<ArrayBuffer>) => {
                         /*  convert Int16Array into Buffer  */
-                        const buf = util.convertI16ToBuf(payload)
-
-                        /*  update chunk  */
-                        chunk.payload = buf
+                        const buf = util.convertI16ToBuf(payload, self.config.audioLittleEndian)
 
-                        /*  forward updated chunk  */
-                        this.push(chunk)
+                        /*  forward cloned chunk with updated payload  */
+                        const chunkNew = chunk.clone()
+                        chunkNew.payload = buf
+                        this.push(chunkNew)
                         callback()
                     }).catch((err: unknown) => {
                         const error = util.ensureError(err)

package/speechflow-cli/src/speechflow-node-a2a-speex.ts

@@ -1,6 +1,6 @@
 /*
 **  SpeechFlow - Speech Processing Flow Graph
-**  Copyright (c) 2024-2025 Dr. Ralf S. Engelschall <rse@engelschall.com>
+**  Copyright (c) 2024-2026 Dr. Ralf S. Engelschall <rse@engelschall.com>
 **  Licensed under GPL 3.0 <https://spdx.org/licenses/GPL-3.0-only>
 */
 
@@ -53,7 +53,9 @@ export default class SpeechFlowNodeA2ASpeex extends SpeechFlowNode {
         const wasmBinary = await fs.promises.readFile(
             path.join(__dirname, "../node_modules/@sapphi-red/speex-preprocess-wasm/dist/speex.wasm"))
         const speexModule = await loadSpeexModule({
-            wasmBinary: wasmBinary.buffer
+            wasmBinary: wasmBinary.buffer.slice(
+                wasmBinary.byteOffset,
+                wasmBinary.byteOffset + wasmBinary.byteLength)
         })
         this.speexProcessor = new SpeexPreprocessor(
             speexModule, this.sampleSize, this.config.audioSampleRate)
@@ -79,7 +81,7 @@ export default class SpeechFlowNodeA2ASpeex extends SpeechFlowNode {
                     callback(new Error("invalid chunk payload type"))
                 else {
                     /*  convert Buffer into Int16Array  */
-                    const payload = util.convertBufToI16(chunk.payload)
+                    const payload = util.convertBufToI16(chunk.payload, self.config.audioLittleEndian)
 
                     /*  process Int16Array in necessary fixed-size segments  */
                     util.processInt16ArrayInSegments(payload, self.sampleSize, (segment) => {
@@ -94,13 +96,12 @@ export default class SpeechFlowNodeA2ASpeex extends SpeechFlowNode {
                             throw new Error("stream already destroyed")
 
                         /*  convert Int16Array back into Buffer  */
-                        const buf = util.convertI16ToBuf(payload)
+                        const buf = util.convertI16ToBuf(payload, self.config.audioLittleEndian)
 
-                        /*  update chunk  */
-                        chunk.payload = buf
-
-                        /*  forward updated chunk  */
-                        this.push(chunk)
+                        /*  forward cloned chunk with updated payload  */
+                        const chunkNew = chunk.clone()
+                        chunkNew.payload = buf
+                        this.push(chunkNew)
                         callback()
                     }).catch((err: unknown) => {
                         const error = util.ensureError(err)

package/speechflow-cli/src/speechflow-node-a2a-vad.ts

@@ -1,6 +1,6 @@
 /*
 **  SpeechFlow - Speech Processing Flow Graph
-**  Copyright (c) 2024-2025 Dr. Ralf S. Engelschall <rse@engelschall.com>
+**  Copyright (c) 2024-2026 Dr. Ralf S. Engelschall <rse@engelschall.com>
 **  Licensed under GPL 3.0 <https://spdx.org/licenses/GPL-3.0-only>
 */
 
@@ -258,6 +258,9 @@ export default class SpeechFlowNodeA2AVAD extends SpeechFlowNode {
                         return
                     }
 
+                    /*  await forthcoming audio chunks (forward declaration)  */
+                    let awaitForthcomingChunks: () => void = () => {}
+
                     /*  flush pending audio chunks  */
                     const flushPendingChunks = () => {
                         let pushed = 0
@@ -289,22 +292,22 @@ export default class SpeechFlowNodeA2AVAD extends SpeechFlowNode {
                                 this.push(chunk)
                                 pushed++
                             }
-                            else if (self.params.mode === "unplugged" && pushed === 0) {
-                                /*  we have to await chunks now, as in unplugged
-                                    mode we else would be never called again until
-                                    we at least once push a new chunk as the result  */
-                                setTimeout(() => {
-                                    if (self.closing || self.queue === null)
-                                        return
-                                    tryToRead()
-                                }, 0)
-                                return
-                            }
+                        }
+
+                        /*  in unplugged mode, if no chunk was pushed (all were
+                            non-speech), we need to wait event-driven for new
+                            data, as the stream won't call read() again until
+                            we push something  */
+                        if (pushed === 0
+                            && !self.closing
+                            && !self.activeEventListeners.has(awaitForthcomingChunks)) {
+                            self.queue.once("write", awaitForthcomingChunks)
+                            self.activeEventListeners.add(awaitForthcomingChunks)
                         }
                     }
 
                     /*  await forthcoming audio chunks  */
-                    const awaitForthcomingChunks = () => {
+                    awaitForthcomingChunks = () => {
                         self.activeEventListeners.delete(awaitForthcomingChunks)
                         if (self.closing)
                             return
@@ -339,16 +342,28 @@ export default class SpeechFlowNodeA2AVAD extends SpeechFlowNode {
 
     /*  close node  */
     async close () {
-        /*  indicate closing  */
-        this.closing = true
-
         /*  cleanup tail timer  */
         if (this.tailTimer !== null) {
             clearTimeout(this.tailTimer)
             this.tailTimer = null
         }
 
-        /*  remove all event listeners  */
+        /*  flush VAD (before closing, as flush triggers callbacks which need active state)  */
+        if (this.vad !== null) {
+            try {
+                const flushPromise = this.vad.flush()
+                const timeoutPromise = new Promise((resolve) => { setTimeout(resolve, 5000) })
+                await Promise.race([ flushPromise, timeoutPromise ])
+            }
+            catch (error) {
+                this.log("warning", `VAD flush error during close: ${error}`)
+            }
+        }
+
+        /*  indicate closing  */
+        this.closing = true
+
+        /*  remove all remaining event listeners  */
         this.activeEventListeners.forEach((listener) => {
             this.queue.removeListener("write", listener)
         })
@@ -360,23 +375,15 @@ export default class SpeechFlowNodeA2AVAD extends SpeechFlowNode {
             this.stream = null
         }
 
-        /*  cleanup queue pointers before closing VAD to prevent callback access  */
-        this.queue.pointerDelete("recv")
-        this.queue.pointerDelete("vad")
-        this.queue.pointerDelete("send")
-
-        /*  close VAD  */
+        /*  destroy VAD  */
         if (this.vad !== null) {
-            try {
-                const flushPromise = this.vad.flush()
-                const timeoutPromise = new Promise((resolve) => { setTimeout(resolve, 5000) })
-                await Promise.race([ flushPromise, timeoutPromise ])
-            }
-            catch (error) {
-                this.log("warning", `VAD flush error during close: ${error}`)
-            }
             this.vad.destroy()
             this.vad = null
         }
+
+        /*  cleanup queue pointers  */
+        this.queue.pointerDelete("recv")
+        this.queue.pointerDelete("vad")
+        this.queue.pointerDelete("send")
     }
 }

package/speechflow-cli/src/speechflow-node-a2a-wav.ts

@@ -1,6 +1,6 @@
 /*
 **  SpeechFlow - Speech Processing Flow Graph
-**  Copyright (c) 2024-2025 Dr. Ralf S. Engelschall <rse@engelschall.com>
+**  Copyright (c) 2024-2026 Dr. Ralf S. Engelschall <rse@engelschall.com>
 **  Licensed under GPL 3.0 <https://spdx.org/licenses/GPL-3.0-only>
 */
 
@@ -183,9 +183,10 @@ export default class SpeechFlowNodeA2AWAV extends SpeechFlowNode {
                             callback(new Error(`WAV not based on ${self.config.audioChannels} channel(s)`))
                             return
                         }
-                        chunk.payload = chunk.payload.subarray(44)
-                        this.push(chunk)
-                        totalSize += chunk.payload.byteLength
+                        const chunkNew = chunk.clone()
+                        chunkNew.payload = chunk.payload.subarray(44)
+                        this.push(chunkNew)
+                        totalSize += chunkNew.payload.byteLength
                         callback()
                     }
                     else {
@@ -210,7 +211,7 @@ export default class SpeechFlowNodeA2AWAV extends SpeechFlowNode {
                         sampleRate:  self.config.audioSampleRate,
                         bitDepth:    self.config.audioBitDepth
                     })
-                    const headerChunk = headerChunkSent?.clone()
+                    const headerChunk = headerChunkSent.clone()
                     headerChunk.payload = headerBuffer
                     headerChunk.meta.set("chunk:seek", 0)
                     this.push(headerChunk)

package/speechflow-cli/src/speechflow-node-a2t-amazon.ts

@@ -1,6 +1,6 @@
 /*
 **  SpeechFlow - Speech Processing Flow Graph
-**  Copyright (c) 2024-2025 Dr. Ralf S. Engelschall <rse@engelschall.com>
+**  Copyright (c) 2024-2026 Dr. Ralf S. Engelschall <rse@engelschall.com>
 **  Licensed under GPL 3.0 <https://spdx.org/licenses/GPL-3.0-only>
 */
 
@@ -42,6 +42,7 @@ class AsyncQueue<T> {
             resolve?.({ value: null, done: true })
         }
         this.queue.length = 0
+        this.queue.push(null)
     }
     async * [Symbol.asyncIterator] (): AsyncIterator<T> {
         while (true) {
@@ -71,8 +72,9 @@ export default class SpeechFlowNodeA2TAmazon extends SpeechFlowNode {
     private client:       TranscribeStreamingClient                | null = null
     private clientStream: AsyncIterable<TranscriptResultStream>    | null = null
     private audioQueue:   AsyncQueue<Uint8Array>                   | null = null
+    private queue:        util.AsyncQueue<SpeechFlowChunk | null>  | null = null
+    private clientStreamStarting                                          = false
     private closing                                                       = false
-    private queue:        util.SingleQueue<SpeechFlowChunk | null> | null = null
 
     /*  construct node  */
     constructor (id: string, cfg: { [ id: string ]: any }, opts: { [ id: string ]: any }, args: any[]) {
@@ -110,10 +112,11 @@ export default class SpeechFlowNodeA2TAmazon extends SpeechFlowNode {
             throw new Error("Amazon Transcribe node currently supports PCM-S16LE audio only")
 
         /*  clear destruction flag  */
-        this.closing = false
+        this.closing             = false
+        this.clientStreamStarting = false
 
         /*  create queue for results  */
-        this.queue = new util.SingleQueue<SpeechFlowChunk | null>()
+        this.queue = new util.AsyncQueue<SpeechFlowChunk | null>()
 
         /*  create a store for the meta information  */
         const metastore = new util.TimeStore<Map<string, any>>()
@@ -138,24 +141,31 @@ export default class SpeechFlowNodeA2TAmazon extends SpeechFlowNode {
 
         /*  start streaming  */
         const ensureAudioStreamActive = async () => {
-            if (this.clientStream !== null || this.closing)
+            if (this.clientStream !== null || this.clientStreamStarting || this.closing)
                 return
-            const language: LanguageCode = this.params.language === "de" ? "de-DE" : "en-US"
-            const command = new StartStreamTranscriptionCommand({
-                LanguageCode: language,
-                EnablePartialResultsStabilization: this.params.interim,
-                ...(this.params.interim ? { PartialResultsStability: "low" } : {}),
-                MediaEncoding: "pcm",
-                MediaSampleRateHertz: this.config.audioSampleRate,
-                AudioStream: audioStream,
-            })
-            const response = await this.client!.send(command)
-            const stream = response.TranscriptResultStream
-            if (!stream)
-                throw new Error("no TranscriptResultStream returned")
-            this.clientStream = stream
+            this.clientStreamStarting = true
+            try {
+                const language: LanguageCode = this.params.language === "de" ? "de-DE" : "en-US"
+                const command = new StartStreamTranscriptionCommand({
+                    LanguageCode: language,
+                    EnablePartialResultsStabilization: this.params.interim,
+                    ...(this.params.interim ? { PartialResultsStability: "low" } : {}),
+                    MediaEncoding: "pcm",
+                    MediaSampleRateHertz: this.config.audioSampleRate,
+                    AudioStream: audioStream,
+                })
+                const response = await this.client!.send(command)
+                const stream = response.TranscriptResultStream
+                if (!stream)
+                    throw new Error("no TranscriptResultStream returned")
+                this.clientStream = stream
+            }
+            catch (err) {
+                this.clientStreamStarting = false
+                throw err
+            }
             ;(async () => {
-                for await (const event of stream) {
+                for await (const event of this.clientStream!) {
                     const te = event.TranscriptEvent
                     if (!te?.Transcript?.Results)
                         continue
@@ -194,6 +204,8 @@ export default class SpeechFlowNodeA2TAmazon extends SpeechFlowNode {
                 }
             })().catch((err: unknown) => {
                 this.log("warning", `failed to establish connectivity to Amazon Transcribe: ${util.ensureError(err).message}`)
+                this.clientStream         = null
+                this.clientStreamStarting = false
             })
         }
 
@@ -259,7 +271,7 @@ export default class SpeechFlowNodeA2TAmazon extends SpeechFlowNode {
                         this.push(null)
                     }
                     else {
-                        self.log("debug", `received data (${chunk.payload.length} bytes): "${chunk.payload}"`)
+                        self.log("debug", `received data (${chunk.payload.length} bytes)`)
                         this.push(chunk)
                     }
                 }).catch((error: unknown) => {
@@ -273,7 +285,8 @@ export default class SpeechFlowNodeA2TAmazon extends SpeechFlowNode {
     /*  close node  */
     async close () {
         /*  indicate closing first to stop all async operations  */
-        this.closing = true
+        this.closing              = true
+        this.clientStreamStarting = false
 
         /*  shutdown stream  */
         if (this.stream !== null) {

package/speechflow-cli/src/speechflow-node-a2t-deepgram.ts

@@ -1,6 +1,6 @@
 /*
 **  SpeechFlow - Speech Processing Flow Graph
-**  Copyright (c) 2024-2025 Dr. Ralf S. Engelschall <rse@engelschall.com>
+**  Copyright (c) 2024-2026 Dr. Ralf S. Engelschall <rse@engelschall.com>
 **  Licensed under GPL 3.0 <https://spdx.org/licenses/GPL-3.0-only>
 */
 
@@ -24,7 +24,7 @@ export default class SpeechFlowNodeA2TDeepgram extends SpeechFlowNode {
     private dg:                Deepgram.LiveClient                      | null = null
     private closing                                                            = false
     private connectionTimeout: ReturnType<typeof setTimeout>            | null = null
-    private queue:             util.SingleQueue<SpeechFlowChunk | null> | null = null
+    private queue:             util.AsyncQueue<SpeechFlowChunk | null>  | null = null
 
     /*  construct node  */
     constructor (id: string, cfg: { [ id: string ]: any }, opts: { [ id: string ]: any }, args: any[]) {
@@ -64,7 +64,7 @@ export default class SpeechFlowNodeA2TDeepgram extends SpeechFlowNode {
                         balance += balanceResponse.result.balances[0]?.amount ?? 0
                 }
             }
-            else if (response?.error !== null)
+            else if (response !== null && response.error !== null)
                 this.log("warning", `API error fetching projects: ${response.error}`)
         }
         catch (error) {
@@ -83,7 +83,7 @@ export default class SpeechFlowNodeA2TDeepgram extends SpeechFlowNode {
         this.closing = false
 
         /*  create queue for results  */
-        this.queue = new util.SingleQueue<SpeechFlowChunk | null>()
+        this.queue = new util.AsyncQueue<SpeechFlowChunk | null>()
 
         /*  create a store for the meta information  */
         const metastore = new util.TimeStore<Map<string, any>>()
@@ -145,7 +145,7 @@ export default class SpeechFlowNodeA2TDeepgram extends SpeechFlowNode {
                 { word: string, punctuated_word?: string, start: number, end: number }[]
             const isFinal     = (data.is_final     as boolean) ?? false
             const speechFinal = (data.speech_final as boolean) ?? false
-            const kind = ((interim && isFinal) || (endpointing > 0 && speechFinal)) ? "final" : "intermediate"
+            const kind = (isFinal || (endpointing > 0 && speechFinal)) ? "final" : "intermediate"
             if (text === "")
                 this.log("info", `empty/dummy text received (start: ${data.start}s, duration: ${data.duration.toFixed(2)}s)`)
             else {
@@ -206,6 +206,13 @@ export default class SpeechFlowNodeA2TDeepgram extends SpeechFlowNode {
                 }
                 resolve(true)
             })
+            this.dg!.once(Deepgram.LiveTranscriptionEvents.Error, (err: Error) => {
+                if (this.connectionTimeout !== null) {
+                    clearTimeout(this.connectionTimeout)
+                    this.connectionTimeout = null
+                }
+                reject(err)
+            })
         })
 
         /*  remember opening time to receive time zero offset  */
@@ -234,7 +241,11 @@ export default class SpeechFlowNodeA2TDeepgram extends SpeechFlowNode {
                         if (chunk.meta.size > 0)
                             metastore.store(chunk.timestampStart, chunk.timestampEnd, chunk.meta)
                         try {
-                            self.dg.send(chunk.payload.buffer) /* intentionally discard all time information */
+                            /*  send buffer (and intentionally discard all time information)  */
+                            self.dg.send(chunk.payload.buffer.slice(
+                                chunk.payload.byteOffset,
+                                chunk.payload.byteOffset + chunk.payload.byteLength
+                            ))
                         }
                         catch (error) {
                             callback(util.ensureError(error, "failed to send to Deepgram"))

package/speechflow-cli/src/speechflow-node-a2t-google.ts

@@ -1,6 +1,6 @@
 /*
 **  SpeechFlow - Speech Processing Flow Graph
-**  Copyright (c) 2024-2025 Dr. Ralf S. Engelschall <rse@engelschall.com>
+**  Copyright (c) 2024-2026 Dr. Ralf S. Engelschall <rse@engelschall.com>
 **  Licensed under GPL 3.0 <https://spdx.org/licenses/GPL-3.0-only>
 */
 
@@ -24,7 +24,7 @@ export default class SpeechFlowNodeA2TGoogle extends SpeechFlowNode {
     /*  internal state  */
     private client:          GoogleSpeech.SpeechClient                                   | null = null
     private recognizeStream: ReturnType<GoogleSpeech.SpeechClient["streamingRecognize"]> | null = null
-    private queue:           util.SingleQueue<SpeechFlowChunk | null>                    | null = null
+    private queue:           util.AsyncQueue<SpeechFlowChunk | null>                     | null = null
     private closing                                                                             = false
 
     /*  construct node  */
@@ -63,7 +63,7 @@ export default class SpeechFlowNodeA2TGoogle extends SpeechFlowNode {
         this.closing = false
 
         /*  create queue for results  */
-        this.queue = new util.SingleQueue<SpeechFlowChunk | null>()
+        this.queue = new util.AsyncQueue<SpeechFlowChunk | null>()
 
         /*  create a store for the meta information  */
         const metastore = new util.TimeStore<Map<string, any>>()
@@ -152,7 +152,8 @@ export default class SpeechFlowNodeA2TGoogle extends SpeechFlowNode {
                     /*  fallback: use result timing  */
                     const resultEnd = result.resultEndTime
                     if (resultEnd) {
-                        tsEnd = Duration.fromMillis(
+                        tsStart = Duration.fromMillis(0).plus(this.timeZeroOffset)
+                        tsEnd   = Duration.fromMillis(
                             (Number(resultEnd.seconds ?? 0) * 1000) +
                             (Number(resultEnd.nanos ?? 0) / 1000000)
                         ).plus(this.timeZeroOffset)