Speech Recognition (ASR) and Speech Synthesis (TTS)

OpenAI-compatible POST /v1/audio/transcriptions for speech-to-text (multipart) and POST /v1/audio/speech for text-to-speech (JSON), callable directly with the OpenAI SDK.

Inference Space provides two OpenAI-compatible speech endpoints:

Capability	Endpoint	Request body	Returns	Required scope
Speech recognition (ASR)	`POST /v1/audio/transcriptions`	`multipart/form-data`	JSON text + sentence timestamps	`ai:asr`
Speech synthesis (TTS)	`POST /v1/audio/speech`	`application/json`	Binary audio	`ai:tts`

Authenticate with Authorization: Bearer $INFERENCE_SPACE_API_KEY (keys start with gk_). See Authentication and API Keys.

Speech Recognition (ASR)

Minimal Example

# Global region; China region is https://cn.inf.space (accelerated) or https://global.inf.space (international)
curl "https://ai.inf.space/v1/audio/transcriptions" \
  -H "Authorization: Bearer $INFERENCE_SPACE_API_KEY" \
  -F "file=@speech.wav" \
  -F "language=zh-CN"

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["INFERENCE_SPACE_API_KEY"],
    # Global region; China region is https://cn.inf.space (accelerated) or https://global.inf.space (international)
    base_url="https://ai.inf.space/v1",
    timeout=300,
)

with open("speech.wav", "rb") as f:
    # model is required by the SDK but ignored by the gateway; any value works
    result = client.audio.transcriptions.create(model="asr", file=f)
print(result.text)

import fs from "node:fs";

const form = new FormData();
form.set("language", "zh-CN");
form.append("file", new Blob([fs.readFileSync("speech.wav")]), "speech.wav");

// Global region; China region is https://cn.inf.space (accelerated) or https://global.inf.space (international)
const resp = await fetch("https://ai.inf.space/v1/audio/transcriptions", {
  method: "POST",
  headers: { Authorization: `Bearer ${process.env.INFERENCE_SPACE_API_KEY}` },
  body: form,
});
const data = await resp.json();
console.log(data.text);

Request Fields

Field	Type	Required	Description
`file` / `audio`	file	Yes	Audio file; both field names are equivalent (the OpenAI SDK sends `file`)
`format`	string	No	`pcm` / `wav` / `mp3` / `ogg`. If omitted, inferred from the file extension, then the MIME type, and treated as `wav` if neither helps. An explicit unsupported value returns 400
`language`	string	No	Recognition language, default `zh-CN`
`sample_rate`	integer	No	Sample rate, default `16000` (mainly relevant for `pcm`)
`enable_speaker_info`	boolean	No	When `true`, attempts speaker separation; utterances then include a `speaker` field
`hotwords`	string (JSON array)	No	Hot words that improve recognition of proper nouns, e.g. `["Inference Space", {"word": "gk_", "weight": 5}]`; up to 100 entries, each ≤ 64 characters
`poll_timeout_ms`	integer	No	Maximum time to wait for this recognition, in ms; capped at 20 minutes (`1200000`)
`model`	string	No	Accepted only for OpenAI SDK compatibility; ignored by the gateway

An empty (0-byte) file returns 400.
This is a synchronous endpoint: the connection stays open until recognition finishes. For long audio, raise your client timeout or split the audio into segments first.

Response

{
  "text": "The weather is nice today, let's take a walk in the park.",
  "duration_ms": 3200,
  "utterances": [
    { "text": "The weather is nice today", "start_time": 0, "end_time": 1600 },
    { "text": "let's take a walk in the park", "start_time": 1600, "end_time": 3200 }
  ]
}

Field	Description
`text`	Full transcript
`duration_ms`	Audio duration in milliseconds; this is also the billed ASR usage
`utterances[]`	Sentences; `start_time` / `end_time` are in milliseconds; include `speaker` when `enable_speaker_info` is on

The response contains only the fields above. OpenAI's response_format values (srt / vtt / verbose_json, etc.) are not supported. Build subtitle files yourself from the utterances timestamps.

Speech Synthesis (TTS)

Minimal Example

# Global region; China region is https://cn.inf.space (accelerated) or https://global.inf.space (international)
curl "https://ai.inf.space/v1/audio/speech" \
  -H "Authorization: Bearer $INFERENCE_SPACE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": "Welcome to Inference Space speech synthesis.",
    "voice": "default",
    "response_format": "mp3"
  }' \
  --output out.mp3

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["INFERENCE_SPACE_API_KEY"],
    # Global region; China region is https://cn.inf.space (accelerated) or https://global.inf.space (international)
    base_url="https://ai.inf.space/v1",
)

# model is required by the SDK but ignored by the gateway; any value works
audio = client.audio.speech.create(
    model="tts",
    voice="default",
    input="Welcome to Inference Space speech synthesis.",
    response_format="mp3",
)
audio.write_to_file("out.mp3")

import fs from "node:fs";

// Global region; China region is https://cn.inf.space (accelerated) or https://global.inf.space (international)
const resp = await fetch("https://ai.inf.space/v1/audio/speech", {
  method: "POST",
  headers: {
    Authorization: `Bearer ${process.env.INFERENCE_SPACE_API_KEY}`,
    "Content-Type": "application/json",
  },
  body: JSON.stringify({
    input: "Welcome to Inference Space speech synthesis.",
    voice: "default",
    response_format: "mp3",
  }),
});
if (!resp.ok) throw new Error(JSON.stringify(await resp.json()));
fs.writeFileSync("out.mp3", Buffer.from(await resp.arrayBuffer()));

Request Fields

Field	Type	Required	Description
`input`	string	Yes	Text to synthesize; must not be empty, ≤ 5000 characters. `text` is accepted as an alias
`voice`	string	No	Voice, default `default`. Available voices are listed in the console; OpenAI voice names (such as `alloy`) are not mapped automatically
`speed`	number	No	Speaking rate, `0.5`–`2.0`, default `1.0`
`response_format`	string	No	Output format `mp3` / `wav` / `pcm`, default `mp3`. `format` is accepted as an alias; `opus` / `aac` / `flac` are not supported and return 400
`model`	string	No	Accepted only for OpenAI SDK compatibility; ignored by the gateway

Response

On success the body is raw audio bytes, and Content-Type depends on the format:

Format	`Content-Type`
`mp3` (default)	`audio/mp3`
`wav`	`audio/wav`
`pcm`	`audio/L16`

TTS is billed by the number of input characters.

Check the HTTP status before writing to disk: a successful response is binary audio, and only error responses are JSON. Do not parse a successful response as JSON, and do not write an error JSON body into an audio file.

Errors

Speech error bodies look like { "error": { "message", "type", "detail"? } }: message is a user-facing description, and detail keeps the original English reason for troubleshooting.

HTTP	Scenario	Action
`400`	Missing audio / empty file / unsupported format; TTS text empty or over 5000 characters, `speed` out of range, unsupported format	Fix the request; do not retry
`401` / `403`	Key missing or invalid / key lacks `ai:asr` or `ai:tts`	See Authentication
`429`	Rate limited or insufficient account balance; the `retry-after` header is set	Back off per `retry-after`; top up if the balance is exhausted
`501`	The capability is not enabled for your organization (`type` is `not_configured`)	Ask your administrator to enable it
`502` / `503` / `504`	Speech service temporarily unavailable, busy, or timed out	Retry a limited number of times with backoff

See Error Codes and Error Handling for the full conventions.

Compatibility Notes

POST /v1/transcribe and POST /v1/synthesize are older paths that still work with the same fields (/v1/synthesize accepts only the text / format field names). New integrations should use /v1/audio/*, which works directly with the OpenAI SDK.

Speech Recognition (ASR) and Speech Synthesis (TTS)

On this page