Kotoba on fal: Realtime Speech APIs, Languages and Pricing

fal announced Kotoba speech APIs on October 19, bringing live translation, streaming transcription, Japanese voice generation and file transcription into its model catalogue. For developers building voice products, the useful change is access through fal, although each endpoint has different language support and billing rules.

This is an availability announcement, rather than evidence that every underlying Kotoba model first launched today. The integration gives teams already using fal another way to add speech, but choosing the right endpoint still requires some care.

Live translation works one direction at a time

The realtime translation API uses a WebSocket connection and returns both translated text and audio. Each session handles one language direction. A two-way conversation therefore needs an application design that accounts for both directions.

Japanese, English, Korean and Mandarin Chinese are supported, with some routes still in beta. Japanese-to-Chinese and Korean-to-Chinese are beta, as are Chinese-to-Japanese and Chinese-to-Korean. The other listed directions are stable. Spanish is not listed for this fal translation endpoint.

The API documentation prices translation at $0.035 per minute of session time, equivalent to $2.10 for an hour. Its playground card instead refers to input-audio minutes. That distinction matters for a connection that stays open while nobody speaks: confirm the billing unit before estimating production costs.

Japanese streaming transcription has its own session cost

The streaming speech-to-text endpoint currently targets Japanese. Its API documentation lists $0.39 per hour of session time. This should not be assumed to have the same language coverage or billing behaviour as the file-transcription endpoint.

The stream sends text deltas that the client appends as they arrive. Its default 800-millisecond turn detection measures the absence of new model text, rather than audio silence. Applications that use a completed turn to trigger a reply need to account for that difference and commit the final turn when ending a session.

There is also an operational limit behind the realtime label. The documentation warns that an idle application can need minutes to start its GPU runner. Teams should test startup separately from an already-running session before promising an instant first response.

Voice generation is priced in bytes

The Kotoba TTS v2 WebSocket API generates Japanese speech with preset speakers and streams the output in chunks. It defaults to 24 kHz PCM output, with other supported formats including Opus.

Pricing is $0.015 per 1,000 UTF-8 bytes. Those are not characters or model tokens. Japanese text commonly occupies multiple bytes per character, so a character-count estimate can understate the bill. Measure the encoded request text when building a cost calculator.

File transcription supports more languages, with a short upload limit

The separate file speech-to-text API accepts audio or video up to 120 seconds long. Longer inputs are rejected, which makes direct uploads of a full meeting unsuitable without a separate segmentation workflow.

Its language options include Japanese, English, Korean, Mandarin Chinese, Taiwanese Mandarin, Hong Kong Cantonese and Spanish. The caller can specify the language, with Japanese the documented default; the endpoint does not automatically detect it. Optional word timestamps and beta speaker diarization support richer transcripts, although diarization can vary between runs.

This endpoint costs $0.25 per hour of input audio, billed by the second. Unlike the documented streaming session prices, the relevant duration here is the submitted audio itself.

For a first integration, pick one task and test its actual constraints: language direction, startup delay, billing duration and output format. Keep the fal API key on a server-side proxy for browser applications. A shared provider account makes access simpler, but it does not make four different speech interfaces interchangeable.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI founders and developer-tool teams turn API launches, technical tradeoffs and integration decisions into clear stories that reach the right users.

Book a 30-minute call.