Skip to content

Open source Grok Bot alternative, Bring your own agent and subscription: message Claude Code, Codex, Gemini CLI, Cursor, Copilot and 40+ AI agents as bots on your own computer, from iPhone, Mac, Windows, Linux or terminal. Agent-agnostic and free. Also a Muse and Dots alternative. Built in Rust.

feat(kit): default voice calls to gpt-transcribe

leepokai20ad6910b699
diff --git a/apps/ios/WhatsNew.md b/apps/ios/WhatsNew.md
index 3b098248..2bd477ef 100644
--- a/apps/ios/WhatsNew.md
+++ b/apps/ios/WhatsNew.md
@@ -5,12 +5,12 @@ listing what iPhone users notice. tools/asc-submit.py reads the released version
a version without one gets a generic "Bug fixes and improvements" line. Delete old sections.
-->
-## 2.6.2
+## 2.6.3
### zh-Hant
-- 輸入框旁的 + 按鈕改成和輸入框一樣高,兩者對齊。
+- 語音通話的「語音轉文字」預設改用 OpenAI 最新的 gpt-transcribe,辨識更準,也更認得 bot 的名字。
### en-US
-- The + button next to the message field now matches the field's height.
+- Voice calls now use OpenAI's newest speech-to-text, gpt-transcribe, by default: more accurate, and better at your bots' names.
diff --git a/docs/features/voice-call.md b/docs/features/voice-call.md
index c145ffcf..7c2eba43 100644
--- a/docs/features/voice-call.md
+++ b/docs/features/voice-call.md
@@ -95,8 +95,12 @@ Each cloud engine has two **modes**:
speech-to-text and is sent as an ordinary message; replies are read with its text-to-speech. Both
run on the computer (`voiceTranscribe`, `voiceSpeak`), so the key never leaves it; utterances stop
at 20 s and replies are spoken in clips (OpenAI 600 characters as AAC, Gemini 150 as WAV) so every
- request and answer fits one 1 MiB channel message. Defaults: `gpt-4o-transcribe` +
- `gpt-4o-mini-tts`, `gemini-3.8-flash` + `gemini-3.8-flash-tts`.
+ request and answer fits one 1 MiB channel message. Defaults: `gpt-transcribe` +
+ `gpt-4o-mini-tts`, `gemini-3.8-flash` + `gemini-3.8-flash-tts`. Speech-to-text gets the app's and
+ the bots' names as hints (gpt-transcribe also as `keywords`). gpt-transcribe returns Chinese in
+ Simplified characters whatever the prompt says, so the client turns it into Traditional
+ (`Hans-Hant`) for a reader whose first Chinese language, or region, is Traditional. Only a
+ model the user picks is saved; the default isn't, so a newer default reaches everyone.
Voice settings (`VoiceChatSettingsView`) open from the call bar's gear and from settings: on the
iPhone, **Computers & settings → Voice chat**; on the Mac, **Account → Voice chat** (this Mac's
diff --git a/host/src/voice.rs b/host/src/voice.rs
index 3b9ba836..d884a8a9 100644
--- a/host/src/voice.rs
+++ b/host/src/voice.rs
@@ -273,11 +273,29 @@ fn classify(p: Provider, models: &[Model]) -> (Vec<String>, Vec<String>, Vec<Str
(realtime, transcribe, speech)
}
+/// Names a speaker is likely to say that speech-to-text would otherwise misspell: the app and
+/// the bots on this computer.
+fn spoken_names(store: &Store) -> Vec<String> {
+ let mut names = vec!["Codync".to_owned()];
+ for bot in store.bots().unwrap_or_default() {
+ if !bot.deleted && !names.contains(&bot.config.name) {
+ names.push(bot.config.name);
+ }
+ }
+ names.truncate(50);
+ names
+}
+
+fn names_hint(names: &[String]) -> String {
+ format!("Names that may come up: {}.", names.join(", "))
+}
+
/// Speech to text for one utterance (`audio`: base64 WAV, 16 kHz mono).
pub async fn transcribe(store: &Store, p: Provider, model: &str, audio: &str) -> Result<Value> {
let key = key(store, p)?;
let model = checked(model)?;
let wav = STANDARD.decode(audio).map_err(|_| anyhow!("audio must be base64"))?;
+ let names = spoken_names(store);
let text = match p {
Provider::OpenAi => {
let boundary = format!("codync-{}", uuid::Uuid::new_v4().simple());
@@ -286,10 +304,19 @@ pub async fn transcribe(store: &Store, p: Provider, model: &str, audio: &str) ->
format!("--{boundary}\r\nContent-Disposition: form-data; name=\"model\"\r\n\r\n{model}\r\n").as_bytes(),
);
// Keeps Traditional Chinese from coming back Simplified, and code terms intact.
- body.extend_from_slice(
- format!("--{boundary}\r\nContent-Disposition: form-data; name=\"prompt\"\r\n\r\n{TRANSCRIBE_HINT}\r\n")
- .as_bytes(),
- );
+ let mut field = |name: &str, value: &str| {
+ body.extend_from_slice(
+ format!("--{boundary}\r\nContent-Disposition: form-data; name=\"{name}\"\r\n\r\n{value}\r\n")
+ .as_bytes(),
+ );
+ };
+ field("prompt", &format!("{TRANSCRIBE_HINT} {}", names_hint(&names)));
+ // Only gpt-transcribe takes keywords; the older models reject the field.
+ if model.starts_with("gpt-transcribe") {
+ for name in &names {
+ field("keywords[]", name);
+ }
+ }
body.extend_from_slice(
format!(
"--{boundary}\r\nContent-Disposition: form-data; name=\"file\"; filename=\"speech.wav\"\r\nContent-Type: audio/wav\r\n\r\n"
@@ -307,7 +334,7 @@ pub async fn transcribe(store: &Store, p: Provider, model: &str, audio: &str) ->
}
Provider::Gemini => {
let body = json!({"contents": [{"parts": [
- {"text": format!("Transcribe this audio exactly as spoken. {TRANSCRIBE_HINT} Reply with the transcript only; reply with nothing if there is no speech.")},
+ {"text": format!("Transcribe this audio exactly as spoken. {TRANSCRIBE_HINT} {} Reply with the transcript only; reply with nothing if there is no speech.", names_hint(&names))},
{"inlineData": {"mimeType": "audio/wav", "data": audio}},
]}]});
let v = send(p, gemini_generate(&key, model).json(&body)).await?;
diff --git a/kit/Sources/CodyncKit/Models/Voice.swift b/kit/Sources/CodyncKit/Models/Voice.swift
index d8b064c8..b86cf4ea 100644
--- a/kit/Sources/CodyncKit/Models/Voice.swift
+++ b/kit/Sources/CodyncKit/Models/Voice.swift
@@ -39,7 +39,7 @@ public enum VoiceProvider: String, CaseIterable, Codable, Sendable {
/// Speech-to-text and text-to-speech defaults for the non-realtime mode.
public var transcribeModel: String {
switch self {
- case .openAI: "gpt-4o-transcribe"
+ case .openAI: "gpt-transcribe"
case .gemini: "gemini-3.8-flash"
}
}
diff --git a/kit/Sources/CodyncUI/Call/CallView.swift b/kit/Sources/CodyncUI/Call/CallView.swift
index 191dcb1e..a7a2fb99 100644
--- a/kit/Sources/CodyncUI/Call/CallView.swift
+++ b/kit/Sources/CodyncUI/Call/CallView.swift
@@ -74,7 +74,7 @@ struct CallView: View {
return client
}
engine = CloudSpeechEngine(provider: provider, speech: .init(
- transcribe: { try await client().voiceTranscribe(provider, model: transcribe, wav: $0) },
+ transcribe: { try await VoiceSettings.inReadersScript(client().voiceTranscribe(provider, model: transcribe, wav: $0)) },
speak: { try await client().voiceSpeak(provider, model: speechModel, voice: voice, text: $0) },
send: { [model, botId] in model.send($0, to: botId) }
))
diff --git a/kit/Sources/CodyncUI/Call/VoiceChatSettingsView.swift b/kit/Sources/CodyncUI/Call/VoiceChatSettingsView.swift
index 06b41561..bfa50f23 100644
--- a/kit/Sources/CodyncUI/Call/VoiceChatSettingsView.swift
+++ b/kit/Sources/CodyncUI/Call/VoiceChatSettingsView.swift
@@ -144,9 +144,13 @@ private struct ProviderVoiceSettings: View {
Color.clear.frame(height: 0)
.task { await load() }
.onChange(of: voice) { _, v in UserDefaults.standard.set(v, forKey: VoiceSettings.voiceKey(provider)) }
- .onChange(of: model) { _, m in UserDefaults.standard.set(m, forKey: VoiceSettings.modelKey(provider)) }
- .onChange(of: transcribeModel) { _, m in UserDefaults.standard.set(m, forKey: VoiceSettings.transcribeKey(provider)) }
- .onChange(of: speechModel) { _, m in UserDefaults.standard.set(m, forKey: VoiceSettings.speechKey(provider)) }
+ .onChange(of: model) { _, m in VoiceSettings.save(m, default: provider.model, forKey: VoiceSettings.modelKey(provider)) }
+ .onChange(of: transcribeModel) { _, m in
+ VoiceSettings.save(m, default: provider.transcribeModel, forKey: VoiceSettings.transcribeKey(provider))
+ }
+ .onChange(of: speechModel) { _, m in
+ VoiceSettings.save(m, default: provider.speechModel, forKey: VoiceSettings.speechKey(provider))
+ }
}
/// The key is on the computer: say so, and offer test / replace / remove.
diff --git a/kit/Sources/CodyncUI/Call/VoiceEngine.swift b/kit/Sources/CodyncUI/Call/VoiceEngine.swift
index 7f19b5bb..5c8990ec 100644
--- a/kit/Sources/CodyncUI/Call/VoiceEngine.swift
+++ b/kit/Sources/CodyncUI/Call/VoiceEngine.swift
@@ -63,7 +63,7 @@ enum VoiceSettings {
UserDefaults.standard.string(forKey: modeKey(provider)) == "speech"
}
- static func transcribeKey(_ provider: VoiceProvider) -> String { "callTranscribeModel.\(provider.rawValue)" }
+ static func transcribeKey(_ provider: VoiceProvider) -> String { "callSpeechToText.\(provider.rawValue)" }
static func speechKey(_ provider: VoiceProvider) -> String { "callSpeechModel.\(provider.rawValue)" }
static func transcribeModel(_ provider: VoiceProvider) -> String {
@@ -74,6 +74,26 @@ enum VoiceSettings {
UserDefaults.standard.string(forKey: speechKey(provider)) ?? provider.speechModel
}
+ /// Keeps a model the user picked; the default isn't stored, so a newer default reaches them.
+ static func save(_ model: String, default fallback: String, forKey key: String) {
+ if model.isEmpty || model == fallback {
+ UserDefaults.standard.removeObject(forKey: key)
+ } else {
+ UserDefaults.standard.set(model, forKey: key)
+ }
+ }
+
+ /// gpt-transcribe writes Chinese in Simplified characters whatever it's told: turn it back
+ /// into Traditional for someone who reads Traditional (their first Chinese language, or a
+ /// Traditional region when none is listed).
+ static func inReadersScript(_ text: String) -> String {
+ let chinese = Locale.preferredLanguages.lazy.map(Locale.Language.init(identifier:)).first { $0.languageCode == .chinese }
+ let traditional = if let chinese { chinese.script == .hanTraditional }
+ else { ["TW", "HK", "MO"].contains(Locale.current.region?.identifier ?? "") }
+ guard traditional else { return text }
+ return text.applyingTransform(StringTransform("Hans-Hant"), reverse: false) ?? text
+ }
+
static func voiceKey(_ provider: VoiceProvider) -> String { "callVoice.\(provider.rawValue)" }
static func voice(_ provider: VoiceProvider) -> String {