feat: macOS architecture, cloud ASR/LLM providers, and 6-step iOS onboarding

- Add macOS menu-bar dictation app with local ASR models (SenseVoice/Qwen3),
  global Option hotkey, and bottom overlay
- Add cloud ASR/LLM providers (Anthropic, Volcengine, Bailian, and more) with
  provider logos, model listing, and connection checks
- Add shared 7-day usage stats UI (UsageStatsCluster / SevenDayUsageChart)
- Add iOS onboarding step 6 for polish LLM setup; hide custom-language-model
  diagnostic toggle behind DEBUG
- Unify iOS onboarding tagline with the macOS brand line ("开口即文字。")
- Rewrite README (Chinese-first, product-oriented) and refresh GitHub Pages
This commit is contained in:
Rocky
2026-07-11 19:10:20 +08:00
parent cdf833935a
commit cc8dd1070a
116 changed files with 6659 additions and 2634 deletions
@@ -0,0 +1,434 @@
// BailianRealtimeASRClient.swift
// OSGKeyboard · Shared
//
// Alibaba Cloud Bailian / DashScope realtime ASR over the classic inference
// WebSocket (`/api-ws/v1/inference`). Matches OpenLess' `bailian.rs` wire
// protocol: run-task PCM binary frames finish-task result events.
import Foundation
struct BailianRealtimeASRClient: CloudASRTranscribing {
let apiKey: String
let endpoint: String
let model: String
let vocabularyID: String?
let session: URLSession
/// 100 ms of 16 kHz / 16-bit / mono PCM.
private static let targetChunkBytes = 3_200
private static let startTimeout: TimeInterval = 8
private static let finalTimeout: TimeInterval = 12
private static let sessionTimeout: TimeInterval = startTimeout + finalTimeout + 4
func prepare(dictionary: PersonalDictionary) async throws {}
func transcribe(
samples: [Float],
sampleRate: Int,
locale: Locale,
dictionary: PersonalDictionary
) async throws -> String {
guard !apiKey.isEmpty else { throw CloudASRError.noAPIKey }
guard sampleRate == 16_000 else {
throw CloudASRError.transport("Bailian realtime expects 16 kHz audio")
}
guard !samples.isEmpty else { throw CloudASRError.emptyTranscript }
let url = try resolvedEndpointURL()
let pcm = Self.pcm16Data(samples: samples)
let taskID = UUID().uuidString.replacingOccurrences(of: "-", with: "")
let resolvedModel = model.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty
? CloudASRModelCatalog.alibabaFunASRRealtime
: model.trimmingCharacters(in: .whitespacesAndNewlines)
var request = URLRequest(url: url)
request.timeoutInterval = 8
request.setValue(
"bearer \(apiKey.trimmingCharacters(in: .whitespacesAndNewlines))",
forHTTPHeaderField: "Authorization"
)
let wsTask = session.webSocketTask(with: request)
wsTask.resume()
return try await withThrowingTaskGroup(of: String.self) { group in
let events = BailianEventStream(task: wsTask)
group.addTask {
defer { events.cancel() }
return try await Self.runSession(
taskID: taskID,
model: resolvedModel,
pcm: pcm,
wsTask: wsTask,
events: events
)
}
group.addTask {
try await Task.sleep(nanoseconds: UInt64(Self.sessionTimeout * 1_000_000_000))
events.cancel()
wsTask.cancel(with: .goingAway, reason: nil)
throw CloudASRError.transport("session timed out")
}
guard let result = try await group.next() else {
throw CloudASRError.emptyTranscript
}
group.cancelAll()
return result.trimmingCharacters(in: .whitespacesAndNewlines)
}
}
/// Settings connection probe: handshake to `task-started` only.
///
/// Reaching `task-started` proves endpoint + `Authorization` + model are
/// all valid which is exactly what "validate connection" must check.
/// It deliberately sends NO audio: DashScope realtime rejects a short
/// silent probe with a `task-failed: emptyAudio`, which is a false
/// negative for a connectivity test. A real auth/quota/model failure
/// still arrives as `task-failed` before `task-started` and surfaces.
func probeConnection() async throws {
guard !apiKey.isEmpty else { throw CloudASRError.noAPIKey }
let url = try resolvedEndpointURL()
let taskID = UUID().uuidString.replacingOccurrences(of: "-", with: "")
let resolvedModel = model.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty
? CloudASRModelCatalog.alibabaFunASRRealtime
: model.trimmingCharacters(in: .whitespacesAndNewlines)
var request = URLRequest(url: url)
request.timeoutInterval = 8
request.setValue(
"bearer \(apiKey.trimmingCharacters(in: .whitespacesAndNewlines))",
forHTTPHeaderField: "Authorization"
)
let wsTask = session.webSocketTask(with: request)
wsTask.resume()
try await withThrowingTaskGroup(of: Void.self) { group in
let events = BailianEventStream(task: wsTask)
group.addTask {
defer { events.cancel() }
try await Self.sendText(
Self.runTaskMessage(taskID: taskID, model: resolvedModel, vocabularyID: nil),
task: wsTask
)
try await events.waitForStarted(timeout: Self.startTimeout)
// Politely end the task; the connection is already proven.
try? await Self.sendText(Self.finishTaskMessage(taskID: taskID), task: wsTask)
}
group.addTask {
try await Task.sleep(nanoseconds: UInt64(Self.startTimeout * 1_000_000_000))
events.cancel()
wsTask.cancel(with: .goingAway, reason: nil)
throw CloudASRError.transport("connection probe timed out")
}
_ = try await group.next()
group.cancelAll()
}
}
private static func runSession(
taskID: String,
model: String,
pcm: Data,
wsTask: URLSessionWebSocketTask,
events: BailianEventStream
) async throws -> String {
try await sendText(
runTaskMessage(taskID: taskID, model: model, vocabularyID: nil),
task: wsTask
)
try await events.waitForStarted(timeout: startTimeout)
var offset = 0
while offset < pcm.count {
let end = min(offset + targetChunkBytes, pcm.count)
try await sendBinary(pcm.subdata(in: offset..<end), task: wsTask)
offset = end
}
// Let the server register the final frames before ending the task.
// Sending `finish-task` in the same instant as the last binary frame
// races the server's audio buffering (root cause of `emptyAudio` on
// very short clips).
try? await Task.sleep(nanoseconds: 120_000_000)
try await sendText(finishTaskMessage(taskID: taskID), task: wsTask)
return try await events.waitForFinalText(timeout: finalTimeout)
}
private func resolvedEndpointURL() throws -> URL {
let raw = endpoint.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty
? CloudASRModelCatalog.bailianDefaultEndpoint
: endpoint.trimmingCharacters(in: .whitespacesAndNewlines)
guard let url = URL(string: raw) else { throw CloudASRError.invalidURL }
return url
}
private static func sendText(_ text: String, task: URLSessionWebSocketTask) async throws {
do {
try await task.send(.string(text))
} catch {
throw CloudASRError.transport(error.localizedDescription)
}
}
private static func sendBinary(_ data: Data, task: URLSessionWebSocketTask) async throws {
do {
try await task.send(.data(data))
} catch {
throw CloudASRError.transport(error.localizedDescription)
}
}
private static func pcm16Data(samples: [Float]) -> Data {
var data = Data()
data.reserveCapacity(samples.count * 2)
for sample in samples {
let scaled = sample * 32_767.0
let clipped = Swift.max(-32_768.0, Swift.min(32_767.0, scaled))
var littleEndian = Int16(clipped.rounded()).littleEndian
withUnsafeBytes(of: &littleEndian) { data.append(contentsOf: $0) }
}
return data
}
/// Overlap-aware join to avoid cumulative duplicate text from interim replays.
static func mergeSegments(_ segments: [String]) -> String {
var result = ""
for segment in segments {
if result.isEmpty {
result = segment
continue
}
let resultChars = Array(result)
let segmentChars = Array(segment)
let maxOverlap = min(resultChars.count, segmentChars.count)
var overlap = 0
if maxOverlap >= 2 {
for length in stride(from: maxOverlap, through: 2, by: -1) {
let tail = resultChars.suffix(length)
let head = segmentChars.prefix(length)
if tail.elementsEqual(head) {
overlap = length
break
}
}
}
result.append(contentsOf: segmentChars.dropFirst(overlap))
}
return result
}
static func runTaskMessage(taskID: String, model: String, vocabularyID: String?) -> String {
var parameters: [String: Any] = [
"sample_rate": 16_000,
"format": "pcm",
]
if let vocabularyID = vocabularyID?.trimmingCharacters(in: .whitespacesAndNewlines),
!vocabularyID.isEmpty {
parameters["vocabulary_id"] = vocabularyID
}
let body: [String: Any] = [
"header": [
"action": "run-task",
"task_id": taskID,
"streaming": "duplex",
],
"payload": [
"task_group": "audio",
"task": "asr",
"function": "recognition",
"model": model,
"parameters": parameters,
"input": [:] as [String: Any],
],
]
guard let data = try? JSONSerialization.data(withJSONObject: body),
let json = String(data: data, encoding: .utf8) else {
return "{}"
}
return json
}
static func finishTaskMessage(taskID: String) -> String {
let body: [String: Any] = [
"header": [
"action": "finish-task",
"task_id": taskID,
"streaming": "duplex",
],
"payload": ["input": [:] as [String: Any]],
]
guard let data = try? JSONSerialization.data(withJSONObject: body),
let json = String(data: data, encoding: .utf8) else {
return "{}"
}
return json
}
}
// MARK: - Concurrent read loop
private final class BailianEventStream: @unchecked Sendable {
private let task: URLSessionWebSocketTask
private let lock = NSLock()
private var started = false
private var finalText: String?
private var failure: Error?
private var readTask: Task<Void, Never>?
init(task: URLSessionWebSocketTask) {
self.task = task
readTask = Task { [weak self] in
await self?.readLoop()
}
}
func cancel() {
readTask?.cancel()
task.cancel(with: .goingAway, reason: nil)
}
func waitForStarted(timeout: TimeInterval) async throws {
let deadline = Date().addingTimeInterval(timeout)
while Date() < deadline {
if let failure = snapshotFailure() { throw failure }
if snapshotStarted() { return }
try await Task.sleep(nanoseconds: 20_000_000)
}
cancel()
throw CloudASRError.transport("task-started timed out")
}
func waitForFinalText(timeout: TimeInterval) async throws -> String {
let deadline = Date().addingTimeInterval(timeout)
while Date() < deadline {
if let failure = snapshotFailure() { throw failure }
if let text = snapshotFinalText() { return text }
try await Task.sleep(nanoseconds: 20_000_000)
}
cancel()
throw CloudASRError.transport("final result timed out")
}
private func snapshotStarted() -> Bool {
lock.lock()
defer { lock.unlock() }
return started
}
private func snapshotFinalText() -> String? {
lock.lock()
defer { lock.unlock() }
return finalText
}
private func snapshotFailure() -> Error? {
lock.lock()
defer { lock.unlock() }
return failure
}
private func readLoop() async {
var finalSegments: [Int64: String] = [:]
var partialSegments: [Int64: String] = [:]
var lastResultText = ""
while !Task.isCancelled {
let message: URLSessionWebSocketTask.Message
do {
message = try await task.receive()
} catch {
publishFailure(CloudASRError.transport(error.localizedDescription))
return
}
let text: String
switch message {
case .string(let value):
text = value
case .data(let data):
text = String(data: data, encoding: .utf8) ?? ""
@unknown default:
continue
}
guard !text.isEmpty else { continue }
guard let json = try? JSONSerialization.jsonObject(with: Data(text.utf8)) as? [String: Any],
let header = json["header"] as? [String: Any] else {
continue
}
let event = header["event"] as? String ?? ""
switch event {
case "task-started":
publishStarted()
case "result-generated":
guard let payload = json["payload"] as? [String: Any],
let output = payload["output"] as? [String: Any],
let sentenceObj = output["sentence"] as? [String: Any] else {
continue
}
if sentenceObj["heartbeat"] as? Bool == true { continue }
guard let rawText = sentenceObj["text"] as? String else { continue }
let trimmed = rawText.trimmingCharacters(in: .whitespacesAndNewlines)
guard !trimmed.isEmpty else { continue }
lastResultText = trimmed
let sentenceID = sentenceObj["sentence_id"] as? Int64 ?? 0
let sentenceEndValue = sentenceObj["sentence_end"]
let sentenceEnd = sentenceEndValue as? Bool ?? false
let endTime = sentenceObj["end_time"] as? Int64 ?? 0
let isFinal = sentenceEndValue != nil ? sentenceEnd : endTime > 0
if isFinal {
finalSegments[sentenceID] = trimmed
partialSegments.removeValue(forKey: sentenceID)
} else {
partialSegments[sentenceID] = trimmed
}
case "task-finished":
if finalSegments.isEmpty {
publishFinal(lastResultText)
} else {
let ordered = finalSegments.keys.sorted().compactMap { finalSegments[$0] }
publishFinal(BailianRealtimeASRClient.mergeSegments(ordered))
}
return
case "task-failed":
let message = header["error_message"] as? String ?? "task failed"
publishFailure(CloudASRError.transport(message))
return
default:
break
}
}
}
private func publishStarted() {
lock.lock()
started = true
lock.unlock()
}
private func publishFinal(_ text: String) {
lock.lock()
finalText = text
lock.unlock()
}
private func publishFailure(_ error: Error) {
lock.lock()
failure = error
lock.unlock()
cancel()
}
}
@@ -13,6 +13,28 @@ public protocol CloudASRTranscribing: Sendable {
locale: Locale,
dictionary: PersonalDictionary
) async throws -> String
/// Settings "validate connection" probe. Verifies transport + auth only.
func probeConnection() async throws
}
extension CloudASRTranscribing {
/// Default probe: transcribe ~1 s of near-silence. An empty transcript
/// counts as success HTTP/streaming providers only need to prove that
/// transport + auth work. Providers whose service rejects silent/short
/// audio (e.g. DashScope realtime returns `emptyAudio`) override this.
public func probeConnection() async throws {
do {
_ = try await transcribe(
samples: [Float](repeating: 0.01, count: 16_000),
sampleRate: 16_000,
locale: Locale(identifier: "zh-CN"),
dictionary: .empty
)
} catch CloudASRError.emptyTranscript {
return
}
}
}
public enum CloudASRClientFactory {
@@ -29,11 +51,12 @@ public enum CloudASRClientFactory {
model: asrModel,
session: session
)
case .alibabaVocabulary:
return AlibabaFunASRClient(
case .bailianStreaming:
return BailianRealtimeASRClient(
apiKey: store.asrApiKey,
endpoint: store.asrBaseURL,
model: asrModel,
persistence: store.cloudASRPersistence,
vocabularyID: nil,
session: session
)
case .prompt:
@@ -44,6 +67,22 @@ public enum CloudASRClientFactory {
model: asrModel,
session: session
)
case .openRouterJson:
return PromptCloudASRClient(
providerId: providerId,
baseURL: store.asrBaseURL,
apiKey: store.asrApiKey,
model: asrModel,
session: session,
requestFormat: .openRouterJson
)
case .volcengineStreaming:
return VolcengineCloudASRClient(
apiKey: store.asrApiKey,
endpoint: store.asrBaseURL,
resourceID: asrModel,
session: session
)
case .localFallback:
return UnsupportedCloudASRClient(providerId: providerId)
}
@@ -151,25 +190,14 @@ struct ZhipuCloudASRClient: CloudASRTranscribing {
}
}
// MARK: - Alibaba Fun-ASR Flash (vocabulary_id + context)
// MARK: - Alibaba Fun-ASR Flash (HTTP sync, context text bias)
/// `UserDefaults` is not `Sendable`; we only touch `persistence` on the
/// actor-isolated cloud ASR path, same as the previous `AppGroupStore` holder.
struct AlibabaFunASRClient: CloudASRTranscribing, @unchecked Sendable {
struct AlibabaFunASRClient: CloudASRTranscribing {
let apiKey: String
let model: String
let persistence: UserDefaults
let session: URLSession
func prepare(dictionary: PersonalDictionary) async throws {
_ = try await AlibabaVocabularySync.ensureVocabularyID(
dictionary: dictionary,
apiKey: apiKey,
targetModel: CloudASRModelCatalog.alibabaVocabularyTargetModel,
defaults: persistence,
session: session
)
}
func prepare(dictionary: PersonalDictionary) async throws {}
func transcribe(
samples: [Float],
@@ -179,14 +207,6 @@ struct AlibabaFunASRClient: CloudASRTranscribing, @unchecked Sendable {
) async throws -> String {
guard !apiKey.isEmpty else { throw CloudASRError.noAPIKey }
let vocabularyID = try await AlibabaVocabularySync.ensureVocabularyID(
dictionary: dictionary,
apiKey: apiKey,
targetModel: CloudASRModelCatalog.alibabaVocabularyTargetModel,
defaults: persistence,
session: session
)
let dataURI = PCMSampleWavEncoder.dataURI(samples: samples, sampleRate: sampleRate)
let urlString = CloudASRModelCatalog.alibabaAPIBase + CloudASRModelCatalog.alibabaMultimodalPath
guard let url = URL(string: urlString) else { throw CloudASRError.invalidURL }
@@ -211,13 +231,10 @@ struct AlibabaFunASRClient: CloudASRTranscribing, @unchecked Sendable {
],
])
var parameters: [String: Any] = [
let parameters: [String: Any] = [
"format": "wav",
"sample_rate": "\(sampleRate)",
]
if let vocabularyID, !vocabularyID.isEmpty {
parameters["vocabulary_id"] = vocabularyID
}
let body: [String: Any] = [
"model": model,
@@ -235,11 +252,11 @@ struct AlibabaFunASRClient: CloudASRTranscribing, @unchecked Sendable {
let (data, response) = try await session.data(for: request)
try ZhipuCloudASRClient.validateHTTP(response: response, data: data)
guard let text = Self.parseText(from: data)?.trimmingCharacters(in: .whitespacesAndNewlines),
!text.isEmpty else {
throw CloudASRError.emptyTranscript
if let text = Self.parseText(from: data)?.trimmingCharacters(in: .whitespacesAndNewlines),
!text.isEmpty {
return text
}
return text
return ""
}
private static func parseText(from data: Data) -> String? {
@@ -256,7 +273,13 @@ struct AlibabaFunASRClient: CloudASRTranscribing, @unchecked Sendable {
}
}
// MARK: - Prompt-biased transcription (OpenAI / MiMo / custom)
// MARK: - Prompt-biased transcription (OpenAI / MiMo / Groq / custom)
enum PromptCloudASRRequestFormat: Sendable {
case multipart
/// OpenRouter expects JSON `{ model, input_audio: { data, format } }`.
case openRouterJson
}
struct PromptCloudASRClient: CloudASRTranscribing {
let providerId: String
@@ -264,6 +287,26 @@ struct PromptCloudASRClient: CloudASRTranscribing {
let apiKey: String
let model: String
let session: URLSession
var requestFormat: PromptCloudASRRequestFormat = .multipart
/// Groq / OpenRouter batch uploads cap around 30 s per request.
private static let whisperCompatibleMaxDurationSeconds: TimeInterval = 30
init(
providerId: String,
baseURL: String,
apiKey: String,
model: String,
session: URLSession,
requestFormat: PromptCloudASRRequestFormat = .multipart
) {
self.providerId = providerId
self.baseURL = baseURL
self.apiKey = apiKey
self.model = model
self.session = session
self.requestFormat = requestFormat
}
func prepare(dictionary: PersonalDictionary) async throws {}
@@ -281,6 +324,13 @@ struct PromptCloudASRClient: CloudASRTranscribing {
dictionary: dictionary
)
}
if requestFormat == .openRouterJson {
return try await transcribeOpenRouterJSON(
samples: samples,
sampleRate: sampleRate,
dictionary: dictionary
)
}
return try await transcribeOpenAIStyle(
samples: samples,
sampleRate: sampleRate,
@@ -288,11 +338,21 @@ struct PromptCloudASRClient: CloudASRTranscribing {
)
}
private func enforceWhisperDuration(samples: [Float], sampleRate: Int) throws {
let duration = Double(samples.count) / Double(sampleRate)
guard duration <= Self.whisperCompatibleMaxDurationSeconds else {
throw CloudASRError.audioTooLong
}
}
private func transcribeOpenAIStyle(
samples: [Float],
sampleRate: Int,
dictionary: PersonalDictionary
) async throws -> String {
if providerId == "groq" || providerId == "openai" || providerId == "custom" {
try enforceWhisperDuration(samples: samples, sampleRate: sampleRate)
}
let wav = PCMSampleWavEncoder.encode(samples: samples, sampleRate: sampleRate)
let trimmedBase = baseURL.hasSuffix("/") ? String(baseURL.dropLast()) : baseURL
let urlString = "\(trimmedBase)/audio/transcriptions"
@@ -335,6 +395,45 @@ struct PromptCloudASRClient: CloudASRTranscribing {
return text
}
private func transcribeOpenRouterJSON(
samples: [Float],
sampleRate: Int,
dictionary: PersonalDictionary
) async throws -> String {
try enforceWhisperDuration(samples: samples, sampleRate: sampleRate)
let wav = PCMSampleWavEncoder.encode(samples: samples, sampleRate: sampleRate)
let trimmedBase = baseURL.hasSuffix("/") ? String(baseURL.dropLast()) : baseURL
let urlString = "\(trimmedBase)/audio/transcriptions"
guard let url = URL(string: urlString) else { throw CloudASRError.invalidURL }
var body: [String: Any] = [
"model": model,
"input_audio": [
"data": wav.base64EncodedString(),
"format": "wav",
],
]
let prompt = dictionary.asrPromptBias(maxCharacters: 600)
if !prompt.isEmpty {
body["prompt"] = prompt
}
var request = URLRequest(url: url)
request.httpMethod = "POST"
request.setValue("application/json", forHTTPHeaderField: "Content-Type")
request.setValue("Bearer \(apiKey)", forHTTPHeaderField: "Authorization")
request.httpBody = try JSONSerialization.data(withJSONObject: body)
request.timeoutInterval = 90
let (data, response) = try await session.data(for: request)
try ZhipuCloudASRClient.validateHTTP(response: response, data: data)
guard let text = Self.parseOpenAIText(from: data)?.trimmingCharacters(in: .whitespacesAndNewlines),
!text.isEmpty else {
throw CloudASRError.emptyTranscript
}
return text
}
private func transcribeMiMo(
samples: [Float],
sampleRate: Int,
@@ -0,0 +1,19 @@
// CloudASRConnectionCheck.swift
// OSGKeyboard · Shared
//
// Settings "validate connection" probe shared by iOS and macOS.
import Foundation
public enum CloudASRConnectionCheck {
/// Verifies the active cloud ASR client can connect + authenticate.
///
/// Each backend decides how to probe (see `CloudASRTranscribing`):
/// HTTP/batch providers transcribe a short silence clip and treat an
/// empty transcript as success; DashScope realtime only handshakes to
/// `task-started` (pushing fake audio makes it fail with `emptyAudio`).
public static func validate(store: any ConfigurationStore) async throws {
let client = CloudASRClientFactory.make(store: store)
try await client.probeConnection()
}
}
@@ -0,0 +1,400 @@
// VolcengineCloudASRClient.swift
// OSGKeyboard · Shared
//
// Volcengine SAUC bigmodel ASR client. The service uses a WebSocket with a
// small custom binary frame wrapper; this file keeps that protocol isolated
// from the HTTP-style cloud ASR clients.
import Foundation
struct VolcengineCloudASRClient: CloudASRTranscribing {
let apiKey: String
let endpoint: String
let resourceID: String
let session: URLSession
private static let targetChunkBytes = 6_400 // 200 ms @ 16 kHz, 16-bit, mono.
private static let finalTimeout: TimeInterval = 12
private static let hotwordCap = 80
func prepare(dictionary: PersonalDictionary) async throws {}
func transcribe(
samples: [Float],
sampleRate: Int,
locale: Locale,
dictionary: PersonalDictionary
) async throws -> String {
guard !samples.isEmpty else { throw CloudASRError.emptyTranscript }
let credentials = try VolcengineCredentials.parse(
apiKey: apiKey,
fallbackResourceID: resolvedResourceID
)
let url = try resolvedEndpointURL()
let pcm = Self.pcm16Data(samples: samples)
let connectID = UUID().uuidString
var request = URLRequest(url: url)
request.timeoutInterval = 8
request.setValue(credentials.appID, forHTTPHeaderField: "X-Api-App-Key")
request.setValue(credentials.accessToken, forHTTPHeaderField: "X-Api-Access-Key")
request.setValue(credentials.resourceID, forHTTPHeaderField: "X-Api-Resource-Id")
request.setValue(connectID, forHTTPHeaderField: "X-Api-Connect-Id")
let task = session.webSocketTask(with: request)
task.resume()
defer {
task.cancel(with: .normalClosure, reason: nil)
}
let firstPayload = try Self.firstFramePayload(connectID: connectID, dictionary: dictionary)
try await send(
VolcengineFrame.build(
messageType: .fullClientRequest,
flags: .positiveSequence,
serialization: .json,
payload: firstPayload,
sequence: 1
),
task: task
)
var sequence = 2
var offset = 0
while offset < pcm.count {
let end = min(offset + Self.targetChunkBytes, pcm.count)
try await send(
VolcengineFrame.build(
messageType: .audioOnlyRequest,
flags: .positiveSequence,
serialization: .none,
payload: pcm.subdata(in: offset..<end),
sequence: Int32(sequence)
),
task: task
)
sequence += 1
offset = end
}
try await send(
VolcengineFrame.build(
messageType: .audioOnlyRequest,
flags: .negativeSequence,
serialization: .none,
payload: Data(),
sequence: -Int32(sequence)
),
task: task
)
let text = try await receiveFinalText(task: task)
let trimmed = text.trimmingCharacters(in: .whitespacesAndNewlines)
guard !trimmed.isEmpty else { throw CloudASRError.emptyTranscript }
return trimmed
}
private var resolvedResourceID: String {
resourceID.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty
? CloudASRModelCatalog.volcengineDefaultResourceID
: resourceID.trimmingCharacters(in: .whitespacesAndNewlines)
}
private func resolvedEndpointURL() throws -> URL {
let raw = endpoint.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty
? CloudASRModelCatalog.volcengineEndpoint
: endpoint.trimmingCharacters(in: .whitespacesAndNewlines)
guard let url = URL(string: raw) else { throw CloudASRError.invalidURL }
return url
}
private func send(_ data: Data, task: URLSessionWebSocketTask) async throws {
do {
try await task.send(.data(data))
} catch {
throw CloudASRError.transport(error.localizedDescription)
}
}
private func receiveFinalText(task: URLSessionWebSocketTask) async throws -> String {
try await withThrowingTaskGroup(of: String.self) { group in
group.addTask {
var lastPartial = ""
while true {
let message = try await task.receive()
let data: Data
switch message {
case .data(let payload):
data = payload
case .string(let string):
data = Data(string.utf8)
@unknown default:
continue
}
guard let frame = VolcengineFrame.parse(data) else { continue }
if frame.messageType == .errorMessage {
let body = String(data: frame.payload, encoding: .utf8) ?? ""
let code = frame.errorCode ?? 0
throw CloudASRError.transport("ASR error \(code): \(body)")
}
guard frame.messageType == .fullServerResponse else { continue }
let parsedText = Self.text(from: frame.payload)
if !parsedText.isEmpty {
lastPartial = parsedText
}
if frame.isFinal {
return parsedText.isEmpty ? lastPartial : parsedText
}
}
}
group.addTask {
try await Task.sleep(nanoseconds: UInt64(Self.finalTimeout * 1_000_000_000))
throw CloudASRError.transport("Volcengine final result timed out")
}
let result = try await group.next()!
group.cancelAll()
return result
}
}
private static func firstFramePayload(
connectID: String,
dictionary: PersonalDictionary
) throws -> Data {
var request: [String: Any] = [
"model_name": "bigmodel",
"enable_itn": true,
"enable_punc": true,
"show_utterances": true,
"enable_speaker_info": true,
]
if let context = hotwordContext(dictionary: dictionary) {
request["context"] = context
}
let payload: [String: Any] = [
"user": ["uid": connectID],
"audio": [
"format": "pcm",
"rate": 16_000,
"bits": 16,
"channel": 1,
"codec": "raw",
],
"request": request,
]
return try JSONSerialization.data(withJSONObject: payload)
}
private static func hotwordContext(dictionary: PersonalDictionary) -> String? {
var seen: [String] = []
for word in dictionary.asrHotwords() {
let trimmed = word.trimmingCharacters(in: .whitespacesAndNewlines)
guard !trimmed.isEmpty else { continue }
guard !seen.contains(where: { $0.caseInsensitiveCompare(trimmed) == .orderedSame }) else {
continue
}
seen.append(trimmed)
if seen.count >= hotwordCap { break }
}
guard !seen.isEmpty else { return nil }
let words = seen.map { ["word": $0] }
guard let data = try? JSONSerialization.data(withJSONObject: ["hotwords": words]) else {
return nil
}
return String(data: data, encoding: .utf8)
}
private static func pcm16Data(samples: [Float]) -> Data {
var data = Data()
data.reserveCapacity(samples.count * 2)
for sample in samples {
let scaled = sample * 32_767.0
let clipped = Swift.max(-32_768.0, Swift.min(32_767.0, scaled))
var littleEndian = Int16(clipped.rounded()).littleEndian
withUnsafeBytes(of: &littleEndian) { data.append(contentsOf: $0) }
}
return data
}
private static func text(from payload: Data) -> String {
guard let json = try? JSONSerialization.jsonObject(with: payload) as? [String: Any],
let result = normalizedResult(from: json) else {
return ""
}
if let utterances = result["utterances"] as? [[String: Any]], !utterances.isEmpty {
let pieces = utterances.compactMap { $0["text"] as? String }
let joined = pieces.joined()
if !joined.isEmpty { return joined }
}
return result["text"] as? String ?? ""
}
private static func normalizedResult(from json: [String: Any]) -> [String: Any]? {
if let result = json["result"] as? [String: Any] {
return result
}
if let results = json["result"] as? [[String: Any]] {
return results.first
}
if json["text"] as? String != nil {
return json
}
return nil
}
}
private struct VolcengineCredentials {
let appID: String
let accessToken: String
let resourceID: String
static func parse(apiKey: String, fallbackResourceID: String) throws -> VolcengineCredentials {
let trimmed = apiKey.trimmingCharacters(in: .whitespacesAndNewlines)
guard !trimmed.isEmpty else { throw CloudASRError.noAPIKey }
if let data = trimmed.data(using: .utf8),
let json = try? JSONSerialization.jsonObject(with: data) as? [String: Any] {
let appID = string(json, keys: ["app_id", "appId", "appid"])
let token = string(json, keys: ["access_token", "accessToken", "token"])
let resourceID = string(json, keys: ["resource_id", "resourceId", "resource"])
?? fallbackResourceID
guard let appID, let token, !resourceID.isEmpty else { throw CloudASRError.noAPIKey }
return VolcengineCredentials(appID: appID, accessToken: token, resourceID: resourceID)
}
let separators = CharacterSet(charactersIn: ":\n,")
let parts = trimmed
.components(separatedBy: separators)
.map { $0.trimmingCharacters(in: .whitespacesAndNewlines) }
.filter { !$0.isEmpty }
guard parts.count >= 2 else { throw CloudASRError.noAPIKey }
let resourceID = parts.count >= 3 ? parts[2] : fallbackResourceID
return VolcengineCredentials(appID: parts[0], accessToken: parts[1], resourceID: resourceID)
}
private static func string(_ json: [String: Any], keys: [String]) -> String? {
for key in keys {
if let value = json[key] as? String {
let trimmed = value.trimmingCharacters(in: .whitespacesAndNewlines)
if !trimmed.isEmpty { return trimmed }
}
}
return nil
}
}
private enum VolcengineMessageType: UInt8 {
case fullClientRequest = 0b0001
case audioOnlyRequest = 0b0010
case fullServerResponse = 0b1001
case errorMessage = 0b1111
}
private enum VolcengineFlags: UInt8 {
case none = 0b0000
case positiveSequence = 0b0001
case lastPacket = 0b0010
case negativeSequence = 0b0011
}
private enum VolcengineSerialization: UInt8 {
case none = 0b0000
case json = 0b0001
}
private struct VolcengineFrame {
let messageType: VolcengineMessageType?
let flags: UInt8
let sequence: Int32?
let errorCode: UInt32?
let payload: Data
var isFinal: Bool {
flags == VolcengineFlags.lastPacket.rawValue
|| flags == VolcengineFlags.negativeSequence.rawValue
|| (sequence ?? 0) < 0
}
static func build(
messageType: VolcengineMessageType,
flags: VolcengineFlags,
serialization: VolcengineSerialization,
payload: Data,
sequence: Int32?
) -> Data {
var data = Data()
data.append(0x11)
data.append((messageType.rawValue << 4) | flags.rawValue)
data.append(serialization.rawValue << 4)
data.append(0x00)
if flags == .positiveSequence || flags == .negativeSequence, let sequence {
data.appendBE32(UInt32(bitPattern: sequence))
}
data.appendBE32(UInt32(payload.count))
data.append(payload)
return data
}
static func parse(_ data: Data) -> VolcengineFrame? {
guard data.count >= 8 else { return nil }
let bytes = [UInt8](data)
let headerSize = Int(bytes[0] & 0x0F) * 4
guard headerSize >= 4, data.count >= headerSize + 4 else { return nil }
let typeRaw = (bytes[1] >> 4) & 0x0F
let messageType = VolcengineMessageType(rawValue: typeRaw)
let flags = bytes[1] & 0x0F
let compression = bytes[2] & 0x0F
guard compression == 0 else { return nil }
var offset = headerSize
var sequence: Int32?
if flags == VolcengineFlags.positiveSequence.rawValue
|| flags == VolcengineFlags.negativeSequence.rawValue {
guard let value = data.readBE32(at: offset) else { return nil }
sequence = Int32(bitPattern: value)
offset += 4
}
if messageType == .errorMessage {
guard let code = data.readBE32(at: offset),
let size = data.readBE32(at: offset + 4) else { return nil }
offset += 8
guard data.count >= offset + Int(size) else { return nil }
return VolcengineFrame(
messageType: messageType,
flags: flags,
sequence: sequence,
errorCode: code,
payload: data.subdata(in: offset..<(offset + Int(size)))
)
}
guard let size = data.readBE32(at: offset) else { return nil }
offset += 4
guard data.count >= offset + Int(size) else { return nil }
return VolcengineFrame(
messageType: messageType,
flags: flags,
sequence: sequence,
errorCode: nil,
payload: data.subdata(in: offset..<(offset + Int(size)))
)
}
}
private extension Data {
mutating func appendBE32(_ value: UInt32) {
var bigEndian = value.bigEndian
Swift.withUnsafeBytes(of: &bigEndian) { append(contentsOf: $0) }
}
func readBE32(at offset: Int) -> UInt32? {
guard count >= offset + 4 else { return nil }
return self[offset..<(offset + 4)].reduce(UInt32(0)) { ($0 << 8) | UInt32($1) }
}
}