Voice-to-text dictation app with local (Nvidia Parakeet/Whisper) and cloud models (BYOK). Privacy-first and available cross-platform.
JavaScript aianthropiccross-platformgeminigroqlinuxmacosnvidiaopen-sourceopenaiparakeetspeech-to-texttranscribewhisperwindows
fix(whisper): raise decoder anti-hallucination thresholds on local transcription — fixes #1458
Gabriel Steinb7fdf80bfeed
diff --git a/src/helpers/whisperServer.js b/src/helpers/whisperServer.jsindex ca8e8746f..541a0734a 100644--- a/src/helpers/whisperServer.js+++ b/src/helpers/whisperServer.js@@ -32,6 +32,15 @@ const DEFAULT_WHISPER_THREADS = 4;const MAX_AUTO_WHISPER_THREADS = 12;const MAX_MANUAL_WHISPER_THREADS = 64;const AUTO_THREAD_RATIO = 0.75;+// Decoder anti-hallucination thresholds sent with every /inference request. whisper.cpp's+// defaults (entropy 2.4, logprob -1.0) let a mostly-silent 30s decode window pass the+// repetition check and emit training-data outro boilerplate ("Thank you for watching",+// "Продолжение следует..."). These values cut the hallucinated-tail rate from 2.25% to+// 0.06% over 4,814 real dictations. See #1458.+const INFERENCE_DECODER_FIELDS = Object.freeze({+ entropy_thold: "2.8",+ logprob_thold: "-1.25",+});function clamp(value, min, max) {return Math.max(min, Math.min(max, value));@@ -854,6 +863,14 @@ class WhisperServerManager extends EventEmitter {`${language || "auto"}\r\n`);+ for (const [name, value] of Object.entries(INFERENCE_DECODER_FIELDS)) {+ parts.push(+ `--${boundary}\r\n` ++ `Content-Disposition: form-data; name="${name}"\r\n\r\n` ++ `${value}\r\n`+ );+ }+// Add initial prompt for custom dictionary wordsif (initialPrompt) {parts.push(@@ -1118,6 +1135,7 @@ class WhisperServerManager extends EventEmitter {module.exports = WhisperServerManager;module.exports.buildWhisperServerArgs = buildWhisperServerArgs;+module.exports.INFERENCE_DECODER_FIELDS = INFERENCE_DECODER_FIELDS;module.exports.parseVulkanDevices = parseVulkanDevices;module.exports.resolveVulkanPinAction = resolveVulkanPinAction;module.exports.getVadSignature = getVadSignature;diff --git a/test/helpers/whisperServerInferenceFields.test.js b/test/helpers/whisperServerInferenceFields.test.jsnew file mode 100644index 000000000..ab88cb3b1--- /dev/null+++ b/test/helpers/whisperServerInferenceFields.test.js@@ -0,0 +1,99 @@+const test = require("node:test");+const assert = require("node:assert/strict");+const http = require("node:http");++const WhisperServerManager = require("../../src/helpers/whisperServer");+const { INFERENCE_DECODER_FIELDS } = require("../../src/helpers/whisperServer");++function startCapturingServer() {+ let captured = null;+ const server = http.createServer((req, res) => {+ const chunks = [];+ req.on("data", (chunk) => chunks.push(chunk));+ req.on("end", () => {+ captured = Buffer.concat(chunks).toString("latin1");+ res.writeHead(200, { "Content-Type": "application/json" });+ res.end(JSON.stringify({ text: "ok" }));+ });+ });+ return new Promise((resolve) => {+ server.listen(0, "127.0.0.1", () =>+ resolve({ server, port: server.address().port, getBody: () => captured })+ );+ });+}++function createManager(port) {+ const manager = new WhisperServerManager();+ manager.ready = true;+ manager.hostname = "127.0.0.1";+ manager.port = port;+ manager.canConvert = true;+ manager.process = {};+ manager.modelPath = "/tmp/model.bin";+ manager._convertToWav = async (buffer) => buffer;+ return manager;+}++function fieldValue(body, name) {+ const match = body.match(+ new RegExp(`Content-Disposition: form-data; name="${name}"\\r\\n\\r\\n([^\\r]*)\\r\\n`)+ );+ return match ? match[1] : null;+}++test("pins the measured anti-hallucination thresholds from #1458", () => {+ assert.deepEqual(INFERENCE_DECODER_FIELDS, {+ entropy_thold: "2.8",+ logprob_thold: "-1.25",+ });+});++test("sends the decoder threshold fields alongside the existing inference fields", async (t) => {+ const { server, port, getBody } = await startCapturingServer();+ t.after(() => server.close());++ const manager = createManager(port);+ const result = await manager.transcribe(Buffer.from("audio"), {+ language: "auto",+ initialPrompt: "OpenWhispr",+ });+ assert.equal(result.text, "ok");++ const body = getBody();+ for (const [name, value] of Object.entries(INFERENCE_DECODER_FIELDS)) {+ assert.equal(fieldValue(body, name), value, `expected ${name}=${value} in the request body`);+ }+ assert.equal(fieldValue(body, "language"), "auto");+ assert.equal(fieldValue(body, "prompt"), "OpenWhispr");+ assert.equal(fieldValue(body, "response_format"), "json");++ // Field order: file, language, decoder thresholds, prompt, response_format, closing boundary.+ const positions = ['name="file"', 'name="language"']+ .concat(Object.keys(INFERENCE_DECODER_FIELDS).map((name) => `name="${name}"`))+ .concat(['name="prompt"', 'name="response_format"'])+ .map((marker) => body.indexOf(marker));+ assert.ok(+ positions.every((pos, i) => pos >= 0 && (i === 0 || pos > positions[i - 1])),+ `expected ordered form fields, got positions ${positions.join(", ")}`+ );++ const boundary = body.slice(0, body.indexOf("\r\n"));+ assert.ok(body.endsWith(`${boundary}--\r\n`), "expected the closing boundary to end the body");+});++test("sends the decoder threshold fields even without an initial prompt", async (t) => {+ const { server, port, getBody } = await startCapturingServer();+ t.after(() => server.close());++ const manager = createManager(port);+ await manager.transcribe(Buffer.from("audio"), { language: "en" });++ const body = getBody();+ for (const [name, value] of Object.entries(INFERENCE_DECODER_FIELDS)) {+ assert.equal(fieldValue(body, name), value, `expected ${name}=${value} in the request body`);+ }+ assert.equal(fieldValue(body, "language"), "en");+ assert.equal(fieldValue(body, "prompt"), null);+ assert.equal(fieldValue(body, "response_format"), "json");+});