kugelaudio 0.2.3 → 0.3.0

package/src/client.ts

@@ -3,26 +3,52 @@
  */
 
 import {
-    AuthenticationError,
-    InsufficientCreditsError,
+    ConnectionError,
     KugelAudioError,
-    RateLimitError,
+    ValidationError,
+    classifyHttpError,
+    classifyWsClose,
+    classifyWsFrame,
+    classifyWsHandshakeError,
 } from './errors';
 import type {
     AudioChunk,
     AudioResponse,
+    CreateVoiceOptions,
     GenerateOptions,
     GenerationStats,
     KugelAudioOptions,
     Model,
     StreamCallbacks,
-    Voice,
+    StreamConfig,
+    StreamingSessionCallbacks,
+    UpdateVoiceOptions,
+    VoiceDetail,
+    VoiceListResponse,
+    VoiceReference,
     WordTimestamp
 } from './types';
 import { base64ToArrayBuffer } from './utils';
 import { getWebSocket } from './websocket';
 
-const DEFAULT_API_URL = 'https://api.kugelaudio.com';
+import type { Region } from './types';
+
+const REGION_URLS: Record<Region, string> = {
+  eu: 'https://api.kugelaudio.com',
+  us: 'https://us-api.kugelaudio.com',
+  global: 'https://global-api.kugelaudio.com',
+};
+
+const REGION_PREFIXES = ['eu-', 'us-', 'global-'] as const;
+
+function parseApiKey(apiKey: string): { cleanKey: string; detectedRegion?: Region } {
+  for (const prefix of REGION_PREFIXES) {
+    if (apiKey.startsWith(prefix)) {
+      return { cleanKey: apiKey.slice(prefix.length), detectedRegion: prefix.slice(0, -1) as Region };
+    }
+  }
+  return { cleanKey: apiKey };
+}
 
 /**
  * Create a new WebSocket instance.
@@ -37,6 +63,23 @@ function createWs(url: string): WebSocket {
 /** WebSocket OPEN readyState constant. */
 const WS_OPEN = 1;
 
+let _languageWarningLogged = false;
+
+function warnIfNoLanguage(
+  language: string | undefined,
+  normalize: boolean | undefined
+): void {
+  const normEnabled = normalize === undefined || normalize;
+  if (!language && normEnabled && !_languageWarningLogged) {
+    _languageWarningLogged = true;
+    console.warn(
+      "[KugelAudio] No 'language' set with normalization enabled — the server " +
+        'will auto-detect the language, adding ~60-150ms to TTFA. Set language ' +
+        "(e.g., language: 'en') for optimal latency."
+    );
+  }
+}
+
 /**
  * Models resource for listing TTS models.
  */
@@ -72,52 +115,211 @@ class VoicesResource {
     language?: string;
     includePublic?: boolean;
     limit?: number;
-  }): Promise<Voice[]> {
+    offset?: number;
+  }): Promise<VoiceListResponse> {
     const params = new URLSearchParams();
     if (options?.language) params.set('language', options.language);
     if (options?.includePublic !== undefined) {
       params.set('include_public', String(options.includePublic));
     }
     if (options?.limit) params.set('limit', String(options.limit));
+    if (options?.offset) params.set('offset', String(options.offset));
 
     const query = params.toString();
     const path = query ? `/v1/voices?${query}` : '/v1/voices';
-    const response = await this.client.request<{ voices: any[] }>('GET', path);
+    const response = await this.client.request<{ voices: any[]; total: number; limit: number; offset: number }>('GET', path);
 
-    return response.voices.map((v) => ({
-      id: v.id,
-      name: v.name,
-      description: v.description,
-      category: v.category,
-      sex: v.sex,
-      age: v.age,
-      supportedLanguages: v.supported_languages || [],
-      sampleText: v.sample_text,
-      avatarUrl: v.avatar_url,
-      sampleUrl: v.sample_url,
-      isPublic: v.is_public || false,
-      verified: v.verified || false,
-    }));
+    return {
+      voices: response.voices.map((v) => ({
+        id: v.id,
+        name: v.name,
+        description: v.description,
+        category: v.category,
+        sex: v.sex,
+        age: v.age,
+        supportedLanguages: v.supported_languages || [],
+        sampleText: v.sample_text,
+        avatarUrl: v.avatar_url,
+        sampleUrl: v.sample_url,
+        isPublic: v.is_public || false,
+        verified: v.verified || false,
+      })),
+      total: response.total,
+      limit: response.limit,
+      offset: response.offset,
+    };
   }
 
   /**
    * Get a specific voice by ID.
    */
-  async get(voiceId: number): Promise<Voice> {
+  async get(voiceId: number): Promise<VoiceDetail> {
     const v = await this.client.request<any>('GET', `/v1/voices/${voiceId}`);
+    return this.mapVoiceDetail(v);
+  }
+
+  /**
+   * Create a new voice.
+   */
+  async create(options: CreateVoiceOptions): Promise<VoiceDetail> {
+    const metadata = {
+      name: options.name,
+      sex: options.sex,
+      description: options.description ?? '',
+      category: options.category ?? 'conversational',
+      age: options.age ?? 'middle_age',
+      quality: options.quality ?? 'mid',
+      supported_languages: options.supportedLanguages ?? ['en'],
+      is_public: options.isPublic ?? false,
+      sample_text: options.sampleText ?? '',
+    };
+
+    const formData = new FormData();
+    formData.append(
+      'metadata',
+      new Blob([JSON.stringify(metadata)], { type: 'application/json' }),
+    );
+
+    if (options.referenceFiles) {
+      for (const file of options.referenceFiles) {
+        formData.append('files', file);
+      }
+    }
+
+    const v = await this.client.requestMultipart<any>('POST', '/v1/voices', formData);
+    return this.mapVoiceDetail(v);
+  }
+
+  /**
+   * Update an existing voice. Only provided fields are updated.
+   */
+  async update(voiceId: number, options: UpdateVoiceOptions): Promise<VoiceDetail> {
+    const payload: Record<string, unknown> = {};
+    if (options.name !== undefined) payload.name = options.name;
+    if (options.description !== undefined) payload.description = options.description;
+    if (options.category !== undefined) payload.category = options.category;
+    if (options.age !== undefined) payload.age = options.age;
+    if (options.sex !== undefined) payload.sex = options.sex;
+    if (options.quality !== undefined) payload.quality = options.quality;
+    if (options.supportedLanguages !== undefined) payload.supported_languages = options.supportedLanguages;
+    if (options.isPublic !== undefined) payload.is_public = options.isPublic;
+    if (options.sampleText !== undefined) payload.sample_text = options.sampleText;
+
+    const v = await this.client.request<any>('PATCH', `/v1/voices/${voiceId}`, payload);
+    return this.mapVoiceDetail(v);
+  }
+
+  /**
+   * Delete a voice.
+   */
+  async delete(voiceId: number): Promise<void> {
+    await this.client.request<any>('DELETE', `/v1/voices/${voiceId}`);
+  }
+
+  // -- Reference management --
+
+  /**
+   * List reference audio files for a voice.
+   */
+  async listReferences(voiceId: number): Promise<VoiceReference[]> {
+    const response = await this.client.request<{ references: any[] }>(
+      'GET',
+      `/v1/voices/${voiceId}/references`,
+    );
+    return response.references.map((r) => this.mapVoiceReference(r));
+  }
+
+  /**
+   * Upload a reference audio file to a voice.
+   *
+   * @param voiceId - Voice ID
+   * @param file - Audio file (File in browser, Blob in Node.js)
+   * @param referenceText - Optional transcript of the reference audio
+   */
+  async addReference(
+    voiceId: number,
+    file: File | Blob,
+    referenceText?: string,
+  ): Promise<VoiceReference> {
+    const formData = new FormData();
+    formData.append('file', file);
+    if (referenceText) {
+      formData.append('reference_text', referenceText);
+    }
+
+    const r = await this.client.requestMultipart<any>(
+      'POST',
+      `/v1/voices/${voiceId}/references`,
+      formData,
+    );
+    return this.mapVoiceReference(r);
+  }
+
+  /**
+   * Delete a reference audio file from a voice.
+   */
+  async deleteReference(voiceId: number, referenceId: number): Promise<void> {
+    await this.client.request<any>(
+      'DELETE',
+      `/v1/voices/${voiceId}/references/${referenceId}`,
+    );
+  }
+
+  // -- Publishing --
+
+  /**
+   * Request publication of a voice. Sets it as public and marks it
+   * as pending verification by an admin.
+   */
+  async publish(voiceId: number): Promise<VoiceDetail> {
+    const v = await this.client.request<any>('POST', `/v1/voices/${voiceId}/publish`);
+    return this.mapVoiceDetail(v);
+  }
+
+  // -- Sample generation --
+
+  /**
+   * Trigger sample audio generation for a voice.
+   */
+  async generateSample(voiceId: number): Promise<VoiceDetail> {
+    const v = await this.client.request<any>(
+      'POST',
+      `/v1/voices/${voiceId}/generate-sample`,
+    );
+    return this.mapVoiceDetail(v);
+  }
+
+  // -- Helpers --
+
+  private mapVoiceDetail(v: any): VoiceDetail {
     return {
       id: v.id,
       name: v.name,
-      description: v.description,
-      category: v.category,
-      sex: v.sex,
+      description: v.description ?? '',
+      generativeVoiceDescription: v.generative_voice_description ?? '',
+      supportedLanguages: v.supported_languages ?? [],
+      category: v.category ?? 'cloned',
       age: v.age,
-      supportedLanguages: v.supported_languages || [],
-      sampleText: v.sample_text,
-      avatarUrl: v.avatar_url,
+      sex: v.sex,
+      quality: v.quality ?? 'mid',
+      isPublic: v.is_public ?? false,
+      verified: v.verified ?? false,
+      pendingVerification: v.pending_verification ?? false,
       sampleUrl: v.sample_url,
-      isPublic: v.is_public || false,
-      verified: v.verified || false,
+      avatarUrl: v.avatar_url,
+      sampleText: v.sample_text ?? '',
+    };
+  }
+
+  private mapVoiceReference(r: any): VoiceReference {
+    return {
+      id: r.id,
+      voiceId: r.voice_id,
+      name: r.name ?? '',
+      referenceText: r.reference_text ?? '',
+      s3Path: r.s3_path ?? '',
+      audioUrl: r.audio_url,
+      isGenerated: r.is_generated ?? false,
     };
   }
 }
@@ -135,6 +337,7 @@ class TTSResource {
     reject: (error: Error) => void;
   }> = new Map();
   private requestCounter = 0;
+  private keepaliveTimer: ReturnType<typeof setInterval> | null = null;
 
   constructor(private client: KugelAudio) {}
 
@@ -207,6 +410,67 @@ class TTSResource {
     };
   }
 
+  /**
+   * Stream audio and return a Node.js Readable stream of raw PCM16 binary data.
+   *
+   * **Node.js only** — this method requires the `stream` built-in module and is
+   * intended for server-side integrations such as Vapi custom TTS endpoints,
+   * Express/Fastify handlers, or any pipeline that expects a Node.js `Readable`.
+   *
+   * Compared to manually wiring `onChunk` to a `Readable`, this method avoids
+   * a common race-condition: the stream object is created and returned **before**
+   * any chunks arrive, so the caller can safely pipe or attach listeners before
+   * the first audio byte is pushed.
+   *
+   * @example Vapi custom TTS endpoint
+   * ```typescript
+   * app.post('/synthesize', (req, res) => {
+   *   res.setHeader('Content-Type', 'audio/pcm');
+   *   res.setHeader('Transfer-Encoding', 'chunked');
+   *
+   *   const readable = client.tts.toReadable({
+   *     text: req.body.message.text,
+   *     modelId: 'kugel-1-turbo',
+   *     sampleRate: req.body.message.sampleRate,
+   *     language: 'en',
+   *   });
+   *
+   *   readable.pipe(res);
+   * });
+   * ```
+   *
+   * @param options - TTS generation options (same as `stream()`)
+   * @param reuseConnection - Reuse the pooled WebSocket connection (default: true)
+   * @returns Node.js Readable stream emitting raw PCM16 binary Buffer chunks
+   */
+  // eslint-disable-next-line @typescript-eslint/no-explicit-any
+  toReadable(options: GenerateOptions, reuseConnection = true): any {
+    // Dynamic require keeps browser bundles free of Node.js built-ins.
+    // eslint-disable-next-line @typescript-eslint/no-require-imports
+    const { Readable } = require('stream') as typeof import('stream');
+    const readable = new Readable({ read() {} });
+
+    this.stream(
+      options,
+      {
+        onChunk: (chunk: AudioChunk) => {
+          readable.push(Buffer.from(chunk.audio, 'base64'));
+        },
+        onFinal: () => {
+          readable.push(null);
+        },
+        onError: (error: Error) => {
+          readable.destroy(error);
+        },
+      },
+      reuseConnection
+    ).catch((error: Error) => {
+      readable.destroy(error);
+    });
+
+    return readable;
+  }
+
   /**
    * Build the WebSocket URL with appropriate auth param.
    */
@@ -265,11 +529,20 @@ class TTSResource {
         this.wsConnection = ws;
         this.wsUrl = url;
         this.setupMessageHandler(ws);
+        this.startKeepalive(ws);
         resolve(ws);
       };
 
-      ws.onerror = () => {
-        reject(new KugelAudioError('WebSocket connection error'));
+      ws.onerror = (event: unknown) => {
+        const underlying = (event as { error?: unknown } | null)?.error ?? event;
+        const typed = classifyWsHandshakeError(underlying);
+        reject(
+          typed ??
+            new ConnectionError(
+              `Could not establish KugelAudio WebSocket connection to ${url}. ` +
+                'Check network connectivity.',
+            ),
+        );
       };
     });
   }
@@ -293,7 +566,7 @@ class TTSResource {
         if (!pending) return;
 
         if (data.error) {
-          const error = this.parseError(data.error);
+          const error = this.parseError(data);
           pending.callbacks.onError?.(error);
           this.pendingRequests.delete(requestId);
           pending.reject(error);
@@ -307,7 +580,6 @@ class TTSResource {
             totalSamples: data.total_samples,
             durationMs: data.dur_ms,
             generationMs: data.gen_ms,
-            ttfaMs: data.ttfa_ms,
             rtf: data.rtf,
             error: data.error,
           };
@@ -347,17 +619,25 @@ class TTSResource {
     };
 
     ws.onclose = (event) => {
-      // Clear connection pool
+      // Clear connection pool and keepalive
+      this.stopKeepalive();
       this.wsConnection = null;
       this.wsUrl = null;
 
-      // Reject all pending requests
+      // Reject all pending requests with appropriate error types
       for (const [id, pending] of this.pendingRequests) {
         pending.callbacks.onClose?.();
-        if (event.code === 4001) {
-          pending.reject(new AuthenticationError('Authentication failed'));
-        } else if (event.code === 4003) {
-          pending.reject(new InsufficientCreditsError('Insufficient credits'));
+        // Only surface server-initiated error close codes; normal closes
+        // (1000, 1001) should not reject pending requests with an error.
+        if (
+          event.code === 4001 ||
+          event.code === 4003 ||
+          event.code === 4029 ||
+          event.code === 4500
+        ) {
+          const error = classifyWsClose(event.code, event.reason);
+          pending.callbacks.onError?.(error);
+          pending.reject(error);
         }
         this.pendingRequests.delete(id);
       }
@@ -365,7 +645,9 @@ class TTSResource {
 
     ws.onerror = () => {
       // Reject all pending requests
-      const error = new KugelAudioError('WebSocket connection error');
+      const error = new ConnectionError(
+        'KugelAudio WebSocket connection error. Check network connectivity.',
+      );
       for (const [id, pending] of this.pendingRequests) {
         pending.callbacks.onError?.(error);
         pending.reject(error);
@@ -400,6 +682,7 @@ class TTSResource {
     options: GenerateOptions,
     callbacks: StreamCallbacks
   ): Promise<void> {
+    warnIfNoLanguage(options.language, options.normalize);
     const ws = await this.getConnection();
     const requestId = ++this.requestCounter;
 
@@ -413,11 +696,14 @@ class TTSResource {
         model_id: options.modelId || 'kugel-1-turbo',
         voice_id: options.voiceId,
         cfg_scale: options.cfgScale ?? 2.0,
+        ...(options.temperature !== undefined && { temperature: options.temperature }),
         max_new_tokens: options.maxNewTokens ?? 2048,
         sample_rate: options.sampleRate ?? 24000,
         normalize: options.normalize ?? true,
         ...(options.language && { language: options.language }),
         ...(options.wordTimestamps && { word_timestamps: true }),
+        ...(options.speed !== undefined && { speed: options.speed }),
+        ...(options.projectId !== undefined && { project_id: options.projectId }),
       }));
     });
   }
@@ -429,6 +715,7 @@ class TTSResource {
     options: GenerateOptions,
     callbacks: StreamCallbacks
   ): Promise<void> {
+    warnIfNoLanguage(options.language, options.normalize);
     return new Promise((resolve, reject) => {
       const url = this.buildWsUrl();
       const ws = createWs(url);
@@ -446,6 +733,8 @@ class TTSResource {
           normalize: options.normalize ?? true,
           ...(options.language && { language: options.language }),
           ...(options.wordTimestamps && { word_timestamps: true }),
+          ...(options.speed !== undefined && { speed: options.speed }),
+          ...(options.projectId !== undefined && { project_id: options.projectId }),
         }));
       };
 
@@ -460,7 +749,7 @@ class TTSResource {
           const data = JSON.parse(messageData);
 
           if (data.error) {
-            const error = this.parseError(data.error);
+            const error = this.parseError(data);
             callbacks.onError?.(error);
             ws.close();
             reject(error);
@@ -474,7 +763,6 @@ class TTSResource {
               totalSamples: data.total_samples,
               durationMs: data.dur_ms,
               generationMs: data.gen_ms,
-              ttfaMs: data.ttfa_ms,
               rtf: data.rtf,
               error: data.error,
             };
@@ -513,27 +801,68 @@ class TTSResource {
         }
       };
 
-      ws.onerror = () => {
-        const error = new KugelAudioError('WebSocket connection error');
+      ws.onerror = (event: unknown) => {
+        const underlying = (event as { error?: unknown } | null)?.error ?? event;
+        const error =
+          classifyWsHandshakeError(underlying) ??
+          new ConnectionError(
+            'KugelAudio WebSocket connection error. Check network connectivity.',
+          );
         callbacks.onError?.(error);
         reject(error);
       };
 
       ws.onclose = (event) => {
         callbacks.onClose?.();
-        if (event.code === 4001) {
-          reject(new AuthenticationError('Authentication failed'));
-        } else if (event.code === 4003) {
-          reject(new InsufficientCreditsError('Insufficient credits'));
+        if (
+          event.code === 4001 ||
+          event.code === 4003 ||
+          event.code === 4029 ||
+          event.code === 4500
+        ) {
+          const error = classifyWsClose(event.code, event.reason);
+          callbacks.onError?.(error);
+          reject(error);
         }
       };
     });
   }
 
+  /**
+   * Start periodic keepalive pings on the pooled connection.
+   * Uses the ws package's ping() in Node.js; silently skips in browsers
+   * where WebSocket doesn't expose a ping method.
+   */
+  private startKeepalive(ws: WebSocket): void {
+    this.stopKeepalive();
+    const intervalMs = this.client.keepalivePingInterval;
+    if (intervalMs == null || intervalMs <= 0) return;
+
+    this.keepaliveTimer = setInterval(() => {
+      if (this.wsConnection !== ws || ws.readyState !== WS_OPEN) {
+        this.stopKeepalive();
+        return;
+      }
+      // eslint-disable-next-line @typescript-eslint/no-explicit-any
+      if (typeof (ws as any).ping === 'function') {
+        // eslint-disable-next-line @typescript-eslint/no-explicit-any
+        (ws as any).ping();
+      }
+    }, intervalMs);
+  }
+
+  private stopKeepalive(): void {
+    if (this.keepaliveTimer !== null) {
+      clearInterval(this.keepaliveTimer);
+      this.keepaliveTimer = null;
+    }
+  }
+
   /**
    * Close the pooled WebSocket connection.
    */
   close(): void {
+    this.stopKeepalive();
     if (this.wsConnection) {
       try {
         this.wsConnection.close();
@@ -545,15 +874,43 @@ class TTSResource {
     }
   }
 
-  private parseError(message: string): Error {
-    const lower = message.toLowerCase();
-    if (lower.includes('auth') || lower.includes('unauthorized')) {
-      return new AuthenticationError(message);
-    }
-    if (lower.includes('credit')) {
-      return new InsufficientCreditsError(message);
-    }
-    return new KugelAudioError(message);
+  private parseError(data: { error?: string; error_code?: string; retry_after?: number }): Error {
+    return classifyWsFrame(data);
+  }
+
+  /**
+   * Create a streaming session for LLM integration.
+   *
+   * The session connects to `/ws/tts/stream` and keeps a persistent
+   * connection across multiple {@link StreamingSession.send} calls.
+   * The server auto-chunks text at sentence boundaries — no client-side
+   * flushing required.
+   *
+   * @param config - Session configuration (voice, model, chunking strategy).
+   * @param callbacks - Callbacks for audio chunks and session lifecycle events.
+   * @returns A {@link StreamingSession} instance. Call `.connect()` before sending.
+   *
+   * @example
+   * ```typescript
+   * const session = client.tts.streamingSession(
+   *   { voiceId: 123, autoMode: true, chunkLengthSchedule: [50, 100, 150, 250] },
+   *   { onChunk: (chunk) => playAudio(chunk.audio) },
+   * );
+   *
+   * session.connect();
+   *
+   * for await (const token of llmStream) {
+   *   session.send(token);
+   * }
+   *
+   * await session.close();
+   * ```
+   */
+  streamingSession(
+    config: StreamConfig,
+    callbacks: StreamingSessionCallbacks
+  ): StreamingSession {
+    return new StreamingSession(this.client, config, callbacks);
   }
 
   /**
@@ -574,7 +931,7 @@ class TTSResource {
    *     console.log(`Audio from ${chunk.contextId}`);
    *     playAudio(chunk.audio);
    *   },
-   *   onContextFinal: (contextId) => {
+   *   onContextClosed: (contextId) => {
    *     console.log(`${contextId} finished`);
    *   },
    * });
@@ -625,8 +982,13 @@ class MultiContextSession {
 
   /**
    * Connect to the multi-context WebSocket endpoint.
+   *
+   * The returned promise resolves once the WebSocket is OPEN so callers can
+   * ``await session.connect(callbacks)`` before invoking
+   * {@link createContext} / {@link send}. Pre-open errors reject with the
+   * typed error.
    */
-  connect(callbacks: import('./types').MultiContextCallbacks): void {
+  connect(callbacks: import('./types').MultiContextCallbacks): Promise<void> {
     this.callbacks = callbacks;
 
     const wsUrl = this.client.ttsUrl
@@ -644,12 +1006,9 @@ class MultiContextSession {
 
     const url = `${wsUrl}/ws/tts/multi?${authParam}=${this.client.apiKey}`;
     this.ws = createWs(url);
+    const ws = this.ws;
 
-    this.ws.onopen = () => {
-      // Connection established, ready to create contexts
-    };
-
-    this.ws.onmessage = (event: { data: unknown }) => {
+    ws.onmessage = (event: { data: unknown }) => {
       try {
         // Handle both browser (string) and Node.js (Buffer) message formats
         const messageData = typeof event.data === 'string' 
@@ -690,10 +1049,6 @@ class MultiContextSession {
           this.callbacks.onChunk?.(chunk);
         }
 
-        if (data.is_final) {
-          this.callbacks.onContextFinal?.(data.context_id);
-        }
-
         if (data.context_closed) {
           this.contexts.delete(data.context_id);
           this.callbacks.onContextClosed?.(data.context_id);
@@ -712,20 +1067,51 @@ class MultiContextSession {
       }
     };
 
-    this.ws.onerror = () => {
-      this.callbacks.onError?.(new KugelAudioError('WebSocket connection error'));
-    };
+    return new Promise<void>((resolve, reject) => {
+      let opened = false;
 
-    this.ws.onclose = (event) => {
-      if (event.code === 4001) {
-        this.callbacks.onError?.(new AuthenticationError('Authentication failed'));
-      } else if (event.code === 4003) {
-        this.callbacks.onError?.(new InsufficientCreditsError('Insufficient credits'));
-      }
-      this.ws = null;
-      this.isStarted = false;
-      this.contexts.clear();
-    };
+      ws.onopen = () => {
+        opened = true;
+        resolve();
+      };
+
+      ws.onerror = (event: unknown) => {
+        const underlying = (event as { error?: unknown } | null)?.error ?? event;
+        const err =
+          classifyWsHandshakeError(underlying) ??
+          new ConnectionError(
+            'KugelAudio multi-context WebSocket connection error. ' +
+              'Check network connectivity.',
+          );
+        if (!opened) reject(err);
+        this.callbacks.onError?.(err);
+      };
+
+      ws.onclose = (event) => {
+        let typedErr: KugelAudioError | null = null;
+        if (
+          event.code === 4001 ||
+          event.code === 4003 ||
+          event.code === 4029 ||
+          event.code === 4500
+        ) {
+          typedErr = classifyWsClose(event.code, event.reason);
+          this.callbacks.onError?.(typedErr);
+        }
+        if (!opened) {
+          reject(
+            typedErr ??
+              new ConnectionError(
+                `KugelAudio multi-context WebSocket closed before ready ` +
+                  `(code ${event.code}).`,
+              ),
+          );
+        }
+        this.ws = null;
+        this.isStarted = false;
+        this.contexts.clear();
+      };
+    });
   }
 
   /**
@@ -749,10 +1135,13 @@ class MultiContextSession {
 
     // Include session config on first context
     if (!this.isStarted) {
+      warnIfNoLanguage(this.config.language, this.config.normalize);
       if (this.config.sampleRate) msg.sample_rate = this.config.sampleRate;
       if (this.config.cfgScale) msg.cfg_scale = this.config.cfgScale;
+      if (this.config.temperature !== undefined) msg.temperature = this.config.temperature;
       if (this.config.maxNewTokens) msg.max_new_tokens = this.config.maxNewTokens;
       if (this.config.normalize !== undefined) msg.normalize = this.config.normalize;
+      if (this.config.language) msg.language = this.config.language;
       if (this.config.inactivityTimeout) msg.inactivity_timeout = this.config.inactivityTimeout;
     }
 
@@ -857,6 +1246,336 @@ class MultiContextSession {
   }
 }
 
+/**
+ * Streaming session for LLM integration via `/ws/tts/stream`.
+ *
+ * The server accumulates text across multiple {@link send} calls and
+ * auto-chunks it at sentence boundaries, keeping the KV cache warm between
+ * chunks for natural prosody.  You never need to call `flush` explicitly —
+ * configure {@link StreamConfig.chunkLengthSchedule} or
+ * {@link StreamConfig.autoMode} instead.
+ *
+ * @example
+ * ```typescript
+ * const session = client.tts.streamingSession({
+ *   voiceId: 123,
+ *   autoMode: true,
+ *   chunkLengthSchedule: [50, 100, 150, 250],
+ * }, {
+ *   onChunk: (chunk) => playAudio(chunk.audio),
+ *   onSessionClosed: (totalSecs) => console.log(`Done: ${totalSecs}s`),
+ * });
+ *
+ * session.connect();
+ *
+ * for await (const token of llmStream) {
+ *   session.send(token);
+ * }
+ *
+ * await session.close();
+ * ```
+ */
+class StreamingSession {
+  private ws: WebSocket | null = null;
+  private config: StreamConfig;
+  private callbacks: StreamingSessionCallbacks;
+  private client: KugelAudio;
+  private configSent = false;
+
+  constructor(client: KugelAudio, config: StreamConfig, callbacks: StreamingSessionCallbacks) {
+    this.client = client;
+    this.config = config;
+    this.callbacks = callbacks;
+  }
+
+  /**
+   * Open the WebSocket connection and authenticate.
+   *
+   * The returned promise resolves once the WebSocket is OPEN, so callers can
+   * ``await session.connect()`` and then ``send()`` without racing the
+   * handshake. Pre-open errors (network failure, 4001 unauthorized, …) reject
+   * the promise with the typed error.
+   */
+  connect(): Promise<void> {
+    const wsUrl = this.client.ttsUrl
+      .replace('https://', 'wss://')
+      .replace('http://', 'ws://');
+
+    let authParam: string;
+    if (this.client.isToken) {
+      authParam = 'token';
+    } else if (this.client.isMasterKey) {
+      authParam = 'master_key';
+    } else {
+      authParam = 'api_key';
+    }
+
+    const url = `${wsUrl}/ws/tts/stream?${authParam}=${this.client.apiKey}`;
+    this.ws = createWs(url);
+    const ws = this.ws;
+
+    ws.onmessage = (event: { data: unknown }) => {
+      try {
+        const messageData = typeof event.data === 'string'
+          ? event.data
+          : event.data instanceof Buffer
+            ? event.data.toString()
+            : String(event.data);
+        const data = JSON.parse(messageData);
+
+        if (data.error) {
+          this.callbacks.onError?.(new KugelAudioError(data.error));
+          return;
+        }
+
+        if (data.audio) {
+          const chunk: AudioChunk = {
+            audio: data.audio,
+            encoding: data.enc || 'pcm_s16le',
+            index: data.idx,
+            sampleRate: data.sr,
+            samples: data.samples,
+          };
+          this.callbacks.onChunk?.(chunk);
+        }
+
+        if (data.word_timestamps) {
+          const timestamps = data.word_timestamps.map((w: Record<string, unknown>) => ({
+            word: w.word as string,
+            startMs: w.start_ms as number,
+            endMs: w.end_ms as number,
+            charStart: w.char_start as number,
+            charEnd: w.char_end as number,
+            score: (w.score as number) ?? 1.0,
+          }));
+          this.callbacks.onWordTimestamps?.(timestamps);
+        }
+
+        if (data.chunk_complete) {
+          this.callbacks.onChunkComplete?.(
+            data.chunk_id ?? 0,
+            data.audio_seconds ?? 0,
+            data.gen_ms ?? 0,
+          );
+        }
+
+        if (data.generation_started) {
+          this.callbacks.onGenerationStarted?.(data.chunk_id ?? 0, data.text ?? '');
+        }
+
+        if (data.session_closed) {
+          this.callbacks.onSessionClosed?.(
+            data.total_audio_seconds ?? 0,
+            data.total_text_chunks ?? 0,
+            data.total_audio_chunks ?? 0,
+          );
+        }
+      } catch (e) {
+        console.error('[KugelAudio] Failed to parse streaming session message:', e);
+      }
+    };
+
+    return new Promise<void>((resolve, reject) => {
+      let opened = false;
+
+      ws.onopen = () => {
+        opened = true;
+        resolve();
+      };
+
+      ws.onerror = (event: unknown) => {
+        const underlying = (event as { error?: unknown } | null)?.error ?? event;
+        const err =
+          classifyWsHandshakeError(underlying) ??
+          new ConnectionError(
+            'KugelAudio streaming WebSocket connection error. ' +
+              'Check network connectivity.',
+          );
+        if (!opened) reject(err);
+        this.callbacks.onError?.(err);
+      };
+
+      ws.onclose = (event) => {
+        let typedErr: KugelAudioError | null = null;
+        if (
+          event.code === 4001 ||
+          event.code === 4003 ||
+          event.code === 4029 ||
+          event.code === 4500
+        ) {
+          typedErr = classifyWsClose(event.code, event.reason);
+          this.callbacks.onError?.(typedErr);
+        }
+        if (!opened) {
+          reject(
+            typedErr ??
+              new ConnectionError(
+                `KugelAudio streaming WebSocket closed before ready ` +
+                  `(code ${event.code}).`,
+              ),
+          );
+        }
+        this.ws = null;
+        this.configSent = false;
+      };
+    });
+  }
+
+  /**
+   * Send a text chunk to the server (e.g. one LLM output token).
+   *
+   * The server buffers text across multiple calls and starts generating at
+   * natural sentence boundaries automatically — no need to call `flush`.
+   *
+   * @param text - Raw text or LLM token to append to the server buffer.
+   * @param flush - Force immediate generation of whatever is buffered.
+   *   **Avoid calling this per-sentence from the client.** Doing so bypasses
+   *   the server's semantic chunking, incurs a fresh model prefill cost on
+   *   every flush, and makes latency *worse*, not better.  Let the server
+   *   handle chunking via `chunkLengthSchedule` / `autoMode` instead.
+   */
+  send(text: string, flush = false): void {
+    if (!this.ws || this.ws.readyState !== WS_OPEN) {
+      throw new KugelAudioError('StreamingSession not connected. Call connect() first.');
+    }
+
+    const msg: Record<string, unknown> = { text, flush };
+
+    if (!this.configSent) {
+      if (this.config.voiceId !== undefined) msg.voice_id = this.config.voiceId;
+      if (this.config.modelId !== undefined) msg.model_id = this.config.modelId;
+      if (this.config.cfgScale !== undefined) msg.cfg_scale = this.config.cfgScale;
+      if (this.config.temperature !== undefined) msg.temperature = this.config.temperature;
+      if (this.config.maxNewTokens !== undefined) msg.max_new_tokens = this.config.maxNewTokens;
+      if (this.config.sampleRate !== undefined) msg.sample_rate = this.config.sampleRate;
+      if (this.config.flushTimeoutMs !== undefined) msg.flush_timeout_ms = this.config.flushTimeoutMs;
+      if (this.config.maxBufferLength !== undefined) msg.max_buffer_length = this.config.maxBufferLength;
+      if (this.config.normalize !== undefined) msg.normalize = this.config.normalize;
+      if (this.config.language !== undefined) msg.language = this.config.language;
+      if (this.config.wordTimestamps) msg.word_timestamps = true;
+      if (this.config.autoMode !== undefined) msg.auto_mode = this.config.autoMode;
+      if (this.config.chunkLengthSchedule?.length) msg.chunk_length_schedule = this.config.chunkLengthSchedule;
+      if (this.config.speed !== undefined) msg.speed = this.config.speed;
+      this.configSent = true;
+    }
+
+    this.ws.send(JSON.stringify(msg));
+  }
+
+  /**
+   * End the current session but keep the WebSocket connection open.
+   *
+   * This allows starting a new session on the same connection, avoiding
+   * the overhead of a new WebSocket handshake (~200-300ms). After calling
+   * this, optionally call {@link updateConfig} to change voice/model settings,
+   * then call {@link send} to start the next session.
+   *
+   * The returned promise resolves once the server confirms with a
+   * `session_closed` message, or after a 15 s **quiet** timeout — i.e. 15 s
+   * elapse without *any* server message arriving. The timer resets on every
+   * incoming frame so a long final flush that streams audio for tens of
+   * seconds is not truncated; only a genuinely silent server trips the fuse.
+   */
+  endSession(): Promise<void> {
+    if (!this.ws || this.ws.readyState !== WS_OPEN) return Promise.resolve();
+
+    const ws = this.ws;
+    // Quiet timeout: resets on every incoming server message. Trips only when
+    // the server has been silent for this long. The previous wall-clock fuse
+    // (10 s total) silently truncated audio when the final flushed chunk
+    // took longer to generate than the budget — see fix in this commit.
+    const QUIET_TIMEOUT_MS = 15_000;
+
+    return new Promise<void>((resolve) => {
+      let settled = false;
+      let timer: ReturnType<typeof setTimeout>;
+
+      const prevMessage = ws.onmessage;
+      const prevClose = ws.onclose;
+
+      const done = () => {
+        if (settled) return;
+        settled = true;
+        clearTimeout(timer);
+        // Restore the original handlers so subsequent endSession() calls
+        // don't stack wrappers and so the typed-error onclose installed
+        // by connect() remains in effect for the next session.
+        ws.onmessage = prevMessage;
+        ws.onclose = prevClose;
+        this.configSent = false;
+        resolve();
+      };
+
+      const armQuietTimer = () => {
+        clearTimeout(timer);
+        timer = setTimeout(done, QUIET_TIMEOUT_MS);
+      };
+
+      armQuietTimer();
+
+      ws.onmessage = (event: MessageEvent) => {
+        // Reset the quiet timer on EVERY incoming frame — audio chunks for
+        // the final flush count as liveness, not just session_closed.
+        armQuietTimer();
+        if (prevMessage) prevMessage.call(ws, event);
+        try {
+          const raw = typeof event.data === 'string'
+            ? event.data
+            : event.data instanceof Buffer
+              ? event.data.toString()
+              : String(event.data);
+          if (JSON.parse(raw).session_closed) done();
+        } catch { /* ignore parse errors */ }
+      };
+
+      ws.onclose = (event: CloseEvent) => {
+        this.ws = null;
+        if (prevClose) prevClose.call(ws, event);
+        done();
+      };
+
+      ws.send(JSON.stringify({ close: true }));
+    });
+  }
+
+  /**
+   * Update session configuration for the next session.
+   *
+   * Call this after {@link endSession} and before the next {@link send}
+   * to change voice, model, language, or other settings.
+   */
+  updateConfig(config: Partial<StreamConfig>): void {
+    Object.assign(this.config, config);
+    this.configSent = false;
+  }
+
+  /**
+   * Close the session and the WebSocket connection.
+   *
+   * For session reuse without closing the connection, use
+   * {@link endSession} instead.
+   *
+   * The returned promise resolves once the server confirms the close with a
+   * `session_closed` message, or after a 15 s **quiet** timeout (no traffic
+   * from the server in that window). Audio frames from the server-side
+   * final-flush of the still-buffered text are delivered to your callbacks
+   * before this promise resolves, and each frame resets the quiet timer.
+   */
+  async close(): Promise<void> {
+    await this.endSession();
+
+    if (this.ws) {
+      try { this.ws.close(); } catch { /* already closed */ }
+      this.ws = null;
+    }
+  }
+
+  /** Whether the underlying WebSocket is open. */
+  get isConnected(): boolean {
+    return this.ws !== null && this.ws.readyState === WS_OPEN;
+  }
+}
+
 /**
  * KugelAudio API client.
  *
@@ -870,13 +1589,13 @@ class MultiContextSession {
  * // List voices
  * const voices = await client.voices.list();
  *
- * // Generate audio with fast model (1.5B params)
+ * // Generate audio with fast model
  * const audio = await client.tts.generate({
  *   text: 'Hello, world!',
  *   modelId: 'kugel-1-turbo',
  * });
  *
- * // Generate audio with premium model (7B params)
+ * // Generate audio with premium model
  * const audio = await client.tts.generate({
  *   text: 'Hello, world!',
  *   modelId: 'kugel-1',
@@ -891,6 +1610,7 @@ export class KugelAudio {
   private _apiUrl: string;
   private _ttsUrl: string;
   private _timeout: number;
+  private _keepalivePingInterval: number | null;
 
   /** Models resource */
   public readonly models: ModelsResource;
@@ -901,17 +1621,37 @@ export class KugelAudio {
 
   constructor(options: KugelAudioOptions) {
     if (!options.apiKey) {
-      throw new Error('API key is required');
+      throw new ValidationError(
+        'KugelAudio API key is missing. Set the KUGELAUDIO_API_KEY ' +
+          'environment variable or pass { apiKey: ... } to the client. ' +
+          'Get a key at https://app.kugelaudio.com/settings/api-keys.',
+      );
     }
 
-    this._apiKey = options.apiKey;
+    const { cleanKey, detectedRegion } = parseApiKey(options.apiKey);
+    this._apiKey = cleanKey;
     this._isMasterKey = options.isMasterKey || false;
     this._isToken = options.isToken || false;
     this._orgId = options.orgId;
-    this._apiUrl = (options.apiUrl || DEFAULT_API_URL).replace(/\/$/, '');
+
+    if (options.apiUrl) {
+      this._apiUrl = options.apiUrl.replace(/\/$/, '');
+    } else {
+      const effectiveRegion = options.region || detectedRegion || 'eu';
+      if (!(effectiveRegion in REGION_URLS)) {
+        throw new ValidationError(
+          `Invalid region '${effectiveRegion}'. Must be one of: ${Object.keys(REGION_URLS).join(', ')}.`,
+        );
+      }
+      this._apiUrl = REGION_URLS[effectiveRegion as Region];
+    }
+
     // If ttsUrl not specified, use apiUrl (backend proxies to TTS server)
     this._ttsUrl = (options.ttsUrl || this._apiUrl).replace(/\/$/, '');
     this._timeout = options.timeout || 60000;
+    this._keepalivePingInterval = options.keepalivePingInterval !== undefined
+      ? options.keepalivePingInterval
+      : 20000;
 
     this.models = new ModelsResource(this);
     this.voices = new VoicesResource(this);
@@ -965,6 +1705,11 @@ export class KugelAudio {
     return this._ttsUrl;
   }
 
+  /** Get keepalive ping interval in milliseconds, or null if disabled. */
+  get keepalivePingInterval(): number | null {
+    return this._keepalivePingInterval;
+  }
+
   /**
    * Close the client and release resources.
    * This closes any pooled WebSocket connections.
@@ -1027,25 +1772,57 @@ export class KugelAudio {
 
       clearTimeout(timeoutId);
 
-      if (response.status === 401) {
-        throw new AuthenticationError('Invalid API key');
+      if (!response.ok) {
+        const text = await response.text();
+        throw classifyHttpError(response.status, text, response.headers);
       }
-      if (response.status === 403) {
-        throw new InsufficientCreditsError('Access denied');
+
+      return await response.json();
+    } catch (error) {
+      clearTimeout(timeoutId);
+      if (error instanceof KugelAudioError) {
+        throw error;
       }
-      if (response.status === 429) {
-        throw new RateLimitError('Rate limit exceeded');
+      if ((error as Error).name === 'AbortError') {
+        throw new ConnectionError(
+          `Request to ${method} ${path} timed out after ${this._timeout}ms.`,
+        );
       }
+      throw new ConnectionError(
+        `Could not reach KugelAudio at ${url}: ${(error as Error).message}. ` +
+          'Check network connectivity.',
+      );
+    }
+  }
+
+  /**
+   * Make a multipart/form-data request (for file uploads).
+   * @internal Used by VoicesResource for reference file uploads.
+   */
+  async requestMultipart<T>(method: string, path: string, formData: FormData): Promise<T> {
+    const url = `${this._apiUrl}${path}`;
+
+    const headers: Record<string, string> = {
+      'X-API-Key': this._apiKey,
+      'Authorization': `Bearer ${this._apiKey}`,
+    };
+
+    const controller = new AbortController();
+    const timeoutId = setTimeout(() => controller.abort(), this._timeout);
+
+    try {
+      const response = await fetch(url, {
+        method,
+        headers,
+        body: formData,
+        signal: controller.signal,
+      });
+
+      clearTimeout(timeoutId);
+
       if (!response.ok) {
         const text = await response.text();
-        let message = `HTTP ${response.status}`;
-        try {
-          const json = JSON.parse(text);
-          message = json.detail || json.error || message;
-        } catch {
-          message = text || message;
-        }
-        throw new KugelAudioError(message, response.status);
+        throw classifyHttpError(response.status, text, response.headers);
       }
 
       return await response.json();
@@ -1055,9 +1832,14 @@ export class KugelAudio {
         throw error;
       }
       if ((error as Error).name === 'AbortError') {
-        throw new KugelAudioError('Request timed out');
+        throw new ConnectionError(
+          `Request to ${method} ${path} timed out after ${this._timeout}ms.`,
+        );
       }
-      throw new KugelAudioError(`Request failed: ${(error as Error).message}`);
+      throw new ConnectionError(
+        `Could not reach KugelAudio at ${url}: ${(error as Error).message}. ` +
+          'Check network connectivity.',
+      );
     }
   }
 }