Direct answer
A Turkish-to-English voice translator is not validated by a typed phrase test or a language selector. A conversation path may capture audio, recognize Turkish speech, translate it, display or read English output, then switch direction for the next speaker. Each stage can fail differently, and one direction does not prove the reverse direction. To test an offline app, pin the app and model setup, prepare consented Turkish recordings with reviewed transcripts, and evaluate the exact conversation task after required assets are installed and connectivity is disabled. Record recognition errors separately from translation errors, retain interruptions and empty responses, and check whether optional speech output is actually available. Broad statements such as “multilingual” do not prove support for Turkish in every task, voice, or locale. This guide provides a test method; it does not claim that Cove or any named phone currently delivers Turkish-to-English offline conversation or a measured quality level.
Scenario: asking a question and hearing the answer in English
A Turkish-speaking traveler asks a short question, reads the English translation, and may tap a speaker button to play it. The other person then replies in English and expects a Turkish response. The app must preserve turn ownership and language direction while handling microphone interruption and optional voice availability.
Decision
Evaluate the two directions as separate tasks and verify each audio route. Define which meaning-critical items—names, numbers, negation, dates, and a follow-up correction—must survive. Do not infer a conversation feature from a one-way speech demo.
Expected outcome
The report shows which turns completed, where recognition or translation changed meaning, whether playback was available, and what happened after interruption or loss of network. It does not generalize beyond the tested versions, voices, and fixtures.
Evidence to collect
The claim becomes reviewable only when the following evidence is attached to the same artifact and test run:
- Turkish audio with reviewed transcripts
- Direction-specific English acceptance rubric
- Turn, interruption, offline, and recovery records
Task quality needs a versioned corpus and acceptance rules. Fluent sample output is not an accuracy result.
Implementation workflow
- Specify the conversation mode and both directions. Write down whether the feature is push-to-talk, alternating turns, or continuous listening. Test Turkish-to-English and English-to-Turkish independently, with the selected source and target visible to the user. State whether the output is text, synthesized speech, or both; do not describe a text-only result as full speech-to-speech conversation.
- Prepare reviewed Turkish and English fixtures. Create consented recordings from speakers representative of the intended task and have fluent reviewers validate each source transcript and acceptable translation meaning. Include names, numbers, negation, dates, corrections, short acknowledgements, and background conditions. Keep a separate typed baseline so audio recognition can be evaluated without confusing it with translation.
- Record capture, recognition, and translation separately. For each turn preserve the audio fixture identity, what the recognizer or audio model produced, and the final English or Turkish output. Record permission state, capture interruption, empty audio, and model-readiness behavior. If the system translates directly from audio without exposing a transcript, say that the recognition stage is not independently observable rather than inventing a transcript score.
- Verify setup and execute the real task offline. Install the application and its required model assets before testing. Then disable Wi-Fi and cellular data, start from the same app state, and complete both conversation directions including playback if it is part of the promise. Test a cold launch, permission denial, a phone-call interruption, cancellation, and retry. Note every stage that fails or attempts a network fallback.
- Score meaning and turn management. Use an evaluation rubric defined before examining outputs. Check whether the output preserves the question, answer, named entities, numbers, and negation; then score fluency separately. Record incorrect speaker direction, repeated or omitted turns, interim revisions, and whether the user could correct a mistake. Do not hide a failed turn in an average of successful responses.
- State language and playback boundaries clearly. Verify any locale or language capability claim against the exact model, app build, and observed behavior. Check target-language TTS initialization and installed voice data separately from translation. Report the device, software, model revision, input fixtures, and which language directions were tested. Avoid saying “supports Turkish offline” if only English text output or one successful sample was observed.
Keep each transition observable. A failure should identify the stage, artifact, runtime, and recovery action without logging private user content.
Acceptance criteria
- The app mode and both Turkish-English directions are explicitly identified and tested separately.
- Audio, recognized text when available, translated output, and optional playback have separate records.
- The complete post-setup conversation task is run offline, including interruption and recovery.
- Language, quality, and playback claims are bounded to the tested app, model, voice, and fixture set.
Decision table
| Dimension | Choice or evidence | How to interpret it |
|---|---|---|
| Typed translation | Turkish text to English text with fixed examples | Useful baseline for translation only; it does not test microphone capture or speech recognition. |
| One-way voice | Turkish audio to English text or speech | Tests one direction and output mode; the reverse conversation path remains unverified. |
| Two-way conversation | Alternating Turkish and English turns with corrections and interruptions | Tests turn ownership, route switching, recovery, and meaning across an exchange. |
| Offline playback | Translate, then request target-language TTS with connectivity disabled | Speech synthesis depends on a separate engine and installed voice; translation alone is not proof. |
Failure patterns to prevent
- Using typed Turkish as speech evidence
- Calling one direction proof of the reverse direction
- Claiming model-language support from a broad multilingual label
Also prevent silent fallback, unpinned artifacts, missing cancellation, and conclusions that combine unlike configurations. Store unsuccessful runs alongside successful ones.
Minimum reproducibility record
| Layer | Record |
|---|---|
| Device | Manufacturer, model, chipset, RAM class, operating-system build |
| Software | Application version and git commit |
| Model | Family, variant, revision, format, file length, hash, quantization |
| Runtime | Name, revision, requested backend, observed backend evidence |
| Workload | Fixture revision, input hash, prompt hash, output policy |
| Outcome | Completed, failed, cancelled, fallback, and privacy-safe diagnostics |
Release checklist
- ☐ The primary query is answered without an unsupported number.
- ☐ Every artifact and runtime is pinned.
- ☐ The representative task and failure policy are explicit.
- ☐ Lifecycle, cancellation, cleanup, and fallback are tested.
- ☐ User-content and network boundaries are documented.
- ☐ Result wording applies only to the recorded configuration.
- ☐ The page links to raw method or evidence when results are added.
Sources and related evidence
- Cove Travel voice translation
- Cove benchmark method
- Android AudioRecord microphone capture API
- Gemma 4 model card and supported task overview
- ACL BLEU machine-translation evaluation paper
Chinese deployment and troubleshooting content is organized in the .奇连 AI 端侧专题.
Frequently asked questions
Does a multilingual model guarantee Turkish voice translation offline?
No. A broad multilingual description does not establish Turkish quality for every task or audio path. Verify the exact model version, language direction, app input route, and complete offline workflow with reviewed Turkish fixtures.
Can I use a Turkish text translation to prove voice support?
No. Typed text bypasses the microphone, audio capture, and recognition stages. Use recorded speech with an independently reviewed transcript, and compare recognition and translation results separately whenever the implementation exposes them.
Should Turkish-to-English and English-to-Turkish be tested together?
They can share a test plan, but each direction needs its own fixtures and results. Translation quality, recognition behavior, and voice availability can differ by direction. Do not use success in one direction as evidence for the other.
What counts as an offline conversation test?
After required installation and model setup, disable connectivity and complete a real exchange from microphone capture through output for both speakers. Include turn switches, playback if promised, interruptions, cancellation, and recovery. Record any hidden or explicit network dependency.