Talyno — On-device Earnings-Call Transcriber & Translator

Understand every earnings call.

Talyno is an iPhone-first app built for financial professionals who need to follow multilingual earnings calls, investor meetings, and analyst conferences in real time. It captures audio, transcribes on-device with a custom finance-tuned WhisperKit model, translates on-device with Apple's Translation framework, and surfaces a live bilingual transcript — no bots, no servers, no account required, one tap to start.

---

Screenshots

---

Architecture overview

iPhone (SwiftUI)
  └── AudioCaptureService (AVAudioEngine)
        └── WhisperKitTranscriber (on-device CoreML / Neural Engine)
              └── TranslationCoordinator (Apple TranslationSession)
                    └── Live bilingual transcript UI

Storage: SwiftData  (Meeting, TranscriptSegment, Speaker)

Privacy guarantee: No audio, transcript, or translation ever leaves the device. There is no backend, no account, and no network activity beyond optional on-demand model downloads (see STT Models below).

---

Features

Feature	Description
Live recording	One-tap capture; real-time transcription shown as bilingual bubbles
Finance-tuned STT	Custom WhisperKit model fine-tuned on earnings calls, analyst Q&A, and IR meetings — works offline from first launch
On-device translation	Apple Translation framework — privacy-safe, no third-party API
Meeting library	Browse, search, and re-open past calls
Review & playback	Scroll transcript, search segments, play back audio
Export / share	Share transcript as plain text or Markdown

---

Privacy

• 100% on-device processing — offline by design.
• Microphone is the only required permission; audio never leaves your iPhone.
• No account, no login, no analytics SDK.
• The only outbound network requests are optional model downloads that the user explicitly triggers in Settings.

---

STT models

Model	Quality	Source
openai_whisper-tiny	Basic — works offline at first launch	Bundled in app
talyno-whisper-small-finance	Accurate — custom finance fine-tune, downloads on demand (~466 MB)	passpoo/talyno-whisperkit-coreml (HuggingFace)

The talyno-whisper-small-finance model is a LoRA fine-tune of whisper-small, trained in two phases on English (SPGISpeech earnings-call audio, 200h) and Chinese (Google FLEURS cmn_hans_cn real recordings + a TTS-synthesised Traditional-Chinese financial-terminology corpus). It can be downloaded in Settings → AI Models. The app continues to work with the bundled Basic model if the download is skipped.

The PyTorch source weights are published at passpoo/talyno-whisper-small-finance; the WhisperKit CoreML conversion the app downloads is at passpoo/talyno-whisperkit-coreml.

See docs/talyno-whisper-finetune-guide-v6.md for the full training methodology.

---

Supported languages

Speech-to-text (STT) is currently trained for English and Chinese only:

Language	STT	Translation
English / English (UK)	✅	✅
繁體中文 / 中文（简体）	✅	✅

Known gap: the in-app language picker (MeetingLanguageCatalog) currently still lists Deutsch, Español, 日本語, and 한국어 as STT-selectable from an earlier model generation. The talyno-whisper-small-finance model actually hosted today (v6, see the finetune guide) was retrained to EN + ZH only, so STT quality for those four languages is currently degraded / unsupported until the app's language list is narrowed to match. Translation (Apple's on-device Translation framework) is unaffected and covers a broader set of language pairs independently of STT.

On-device translation packs are downloaded the first time a language pair is used (handled automatically by TranslationPackManager).

---

Prerequisites

Tool	Version
Xcode	16+
iOS target	18.0+
XcodeGen	any (brew install xcodegen)

---

Setup

git clone <repo>
cd Talyno/ios

# Generate the Xcode project from project.yml
xcodegen generate

# Open in Xcode
open Talyno.xcodeproj

No secrets, no backend, no environment variables needed — just clone, generate, and run.

---

Build & test

# Build
xcodebuild -scheme Talyno \
  -destination 'platform=iOS Simulator,name=iPhone 17' \
  build

# Run unit tests
xcodebuild -scheme Talyno \
  -destination 'platform=iOS Simulator,name=iPhone 17' \
  test

# Lint
swiftlint lint --strict

---

Contributing

• Branch off main; open a PR with a clear description.
• Discuss architecture changes in the PR before implementation.
• See CLAUDE.md for coding conventions (Swift 6 strict concurrency, design tokens, SwiftData-only persistence).
• See SECURITY.md to report a vulnerability.

---

License

Talyno's source code is licensed under the Apache License 2.0.